Nondestructive Food Evaluation Techniques To Anyaluze Properties and Quality Food Science and Technology

Nondestructive
Food Evaluation
FOOD SCIENCE AND TECHNOLOGY

A Series of Monographs, Textbooks, and Reference Books
EDITORIAL BOARD
Senior Editors
Owen R. Fennema University of Wisconsin-Madison

Marcus Karel Rutgers University (emeritus)
Gary W. Sanderson Universal Foods Corporation (retired)
Pieter Walstra Wageningen Agricultural University
John R. Whitaker University of California-Davis
Additives P. MichaelDavidson University of Tennessee-Knoxville
Dairy science James L. Steele University of Wisconsin-Madison
Flavor chemistry and sensory analysis John Thorngate University of
Idaho-Moscow
Foodengineering Daryl B. Lund ComellUniversity
Health and disease Seppo Salminen University of Turku, Finland
Nutrition and nutraceuticals MarkDreher Mead Johnson Nutritionals
Processingandpreservation Gustavo W. Barbosa-Chnovas Washington
State University-Pullman
Safety and toxicology Sanford Miller University of Texas-Austin
1. Flavor Research: Principles and Techniques, R. Teranishi, 1. Hornstein,

P. Issenberg, and E. L. Wick
2. Principles of Enzymology for the Food Sciences,
John R. Whitaker
3. Low-TemperaturePreservation of Foods and Living Matter, OwenR.
Fennema, William D. Powrie, and ElmerH. Marth
4. Principles of Food Science
Part I: Food Chemistry, edited by OwenR. Fennema
Part It: Physical Methods of Food Preservation, Marcus Karel, Owen R.
Fennema, andD a y / 6. Lund
5. Food Emulsions,edited by Stig E. Friberg
6. Nutritional and Safety Aspects of Food Processing,edited by Steven R.
Tannenbaum
7. Flavor Research: Recent Advances, edited by R. Temnishi, Robert A.
Flath, and Hiroshi Sugisawa
8. Computer-AidedTechniques in FoodTechnology, editedbyIsrael
SWUY
9. Handbook of Tropical Foods,edited by Harvey T. Chan
I O . Antimicrobials in Foods, edited by Alfred Lany Branen and P. Michael
Davidson
11. Food Constituents and Food Residues: Their Chromatographic Determination, edited by James F. Lawrence
editedbyLewis D. Stegink
12.Aspartame:PhysiologyandBiochemistry,
and L. J. Filer, Jr.
13.HandbookofVitamins:Nutritional,Biochemical,andClinicalAspects,
edited by LawrenceJ. Machlin
14. Starch Conversion Technology, edited by G. M. A. van Beynum and J.
A. Roels
15.FoodChemistry:SecondEdition,RevisedandExpanded,
edited by
Owen R. Fennema
16.SensoryEvaluationofFood:StatisticalMethodsandProcedures,
Michael O'Mahony
17. Alternative Sweeteners, edited by Lyn O'Brien Nabors and Robert
C.
Gelardi
18. Citrus Fruits and Their Products: Analysis and Technology,
S. V. Ting
and Russell L. Rouseff
19. Engineering Properties of Foods,edited by M.A. Rao and S. S. H. Rizvi
20. Umami: A Basic Taste, edited by Yojiro Kawamura and Morley R. Kare
21. Food Biotechnology, edited by Dietrich Knorr
22.FoodTexture:InstrumentalandSensoryMeasurement,
editedby
Howard R. Moskowitz
23. Seafoods and Fish Oils in Human Health and Disease,John E. Kinsella
24. Postharvest Physiology of Vegetables, edited by J. Weichmann
25. Handbook of Dietary Fiber: An Applied Approach,Mark L. Dreher
26. Food Toxicology, PartsA and B, Jose M. Concon
27. Modem Carbohydrate Chemistry, Roger W. Binkley
28. Trace Minerals in Foods, edited by Kenneth T. Smith
29. Protein Quality and the Effects of Processing, edited by R. DixonPhillips
and John W. Finley
30. Adulteration of Fruit Juice Beverages, edited by Steven Nagy, John A.
Attaway, and MarthaE. Rhodes
31. Foodborne Bacterial Pathogens, edited by Michael P. Doyle
32. Legumes: Chemistry, Technology, and Human Nutrition, edited by Ruth
H. Manhews
33.IndustrializationofIndigenousFermentedFoods,
editedbyKeith
H.
Steinkraus
34. International Food Regulation Handbook: Policy0 Science 0 Law, edited
by RogerD. Middlekauff and Philippe Shubik
35. Food Additives, edited by A. Lany Branen, P. Michael Davidson, and
Seppo Salminen
36. Safety of Irradiated Foods, J. F. Diehl
37. Omega3 Fatly Acids in Health and Disease, edited by Robert S. Lees
and Marcus Karel
38. FoodEmulsions:SecondEdition,RevisedandExpanded,
editedby
Kire Larsson and StigE. Friberg
39.Seafood:EffectsofTechnologyonNutrition,
George M. Pigonand
Barbee W. Tucker
40. Handbook of Vitamins: Second Edition, Revised and Expanded,
edited
by LawrenceJ. Machlin
41.HandbookofCerealScienceandTechnology,
Klaus J. Lorenzand
Karel Kulp
42. Food Processing Operations and Scale-up,
Kenneth J. Valentas, Leon
Levine, andJ. Peter Clark
43. Fish Quality Control by Computer Vision,
edited by L. F. Pau and R.
Olafsson
44. Volatile Compounds in Foods and Beverages,edited by Henk Maarse
45. Instrumental Methods for Quality Assurance in Foods, edited by Daniel
Y. C. Fung and RichardF. Matthews
46. Listeria, Listeriosis, and Food Safety,Elliot T. Ryser and Elmer H. Marth
47. Acesulfame-K, edited by D. G. Mayer andF. H. Kemper
48. Alternative Sweeteners: Second Edition, Revised and Expanded, edited
by Lyn O'Brien Nabors and Robert C. Gelardi
49. Food Extrusion Science and Technology, edited by Jozef L. Kokini, ChiTang Ho, and MukundV. Karwe
50. Surimi Technology, edited by Tyre C. Lanier and Chong M. Lee
51. Handbook of Food Engineering,edited by Dennis R. Heldman and Daryl
B. Lund
52. Food Analysis by HPLC, edited by Leo M. L. Nollet
53.FattyAcids in FoodsandTheirHealthImplications,
editedbyChing
Kuang Chow
54. Clostridium botulinum: Ecology and Controlin Foods, edited by Andreas
H. W. Hauschild and KarenL. Dodds
55. Cereals in Breadmaking: A Molecular Colloidal Approach,Ann-Charlotte
Eliasson andKire Larsson
56. Low-Calorie Foods Handbook,edited by Aaron M. Altschul
57. Antimicrobials in Foods: Second Edition, Revised and Expanded,edited
by P. Michael Davidson and Alfred Larry Branen
58. Lactic Acid Bacteria, edited by Seppo Salminen and Atte von Wright
59. Rice Science and Technology,edited by Wayne E. Marshall and James
1. Wadsworth
60.FoodBiosensorAnalysis,
editedbyGabrieleWagnerandGeorgeG.
Guilbault
61. Principles of Enzymology for the Food Sciences: Second Edition, John
R. Whifaker
62. Carbohydrate Polyesters as Fat Substitutes, edited by Casimir C. Akoh
and Barry G. Swanson
63.
Engineering
Properties
of
Foods:
Second
Edition,
Revised
and
Expanded, edited by M. A. Rao and S. S. H. Rimi
64. Handbook of Brewing, edited by William A. Hardwick
65.AnalyzingFood
for NutritionLabelingandHazardousContaminants,
edited by IkeJ. Jeon and William G. lkins
66. Ingredient Interactions: Effects on Food Quality, edited by Anilkumar G.
Gaonkar
67.FoodPolysaccharidesandTheirApplications,
editedby Alisfair M.
Stephen
68. Safety of Irradiated Foods: Second Edition, Revised and Expanded,
J.
F. Diehl
69. Nutrition Labeling Handbook, edited by Ralph Shapiro
70.Handbookof Fruit ScienceandTechnology:Production, COmpOSitiOn,

Storage, and Processing,edited by D. K. Salunkhe andS. S. Kadam
71.FoodAntioxidants:Technological,Toxicological,andHealthPerspectives, edited by D. L. Madhavi, S. S. Deshpande, and D. K. Salunkhe
72. Freezing Effects on Food Quality,edited by Lester E. Jeremiah
73.HandbookofIndigenousFermentedFoods:SecondEdition,Revised
and Expanded,edited by KeithH. Steinkraus
74. Carbohydrates in Food, edited by Ann-Charlotte Eliasson
75.BakedGoodsFreshness:Technology,Evaluation,and
Inhibition of
Staling, edited by RonaldE. Hebeda and HenryF. Zobel
76. Food Chemistry: Third Edition, edited by Owen R. Fennema
77.HandbookofFoodAnalysis:Volumes
1 and 2, edited by Leo M. L.
Nollet
78. Computerized Control Systems in the Food Industry, edited by Gauri S.
Mittal
79. Techniques for Analyzing Food Aroma,edited by Ray Marsili
80. Food Proteins and Their Applications, edited by Srinivasan Damodaran
and Alain Paraf
81. Food Emulsions: Third Edition, Revised and Expanded,edited by Stig E.
Friberg and K5re Larsson
82. Nonthermal Preservation of Foods, Gustavo V. Barbosa-Canovas, Usha
R. Pothakamury, Enrique Palou, and BarryG. Swanson
83. Milk and Dairy Product Technology, Edgar Spreer
84.AppliedDairyMicrobiology,
editedby /mer H. MarthandJames L.
Steele
85.LacticAcidBacteria:MicrobiologyandFunctionalAspects:Second
Edition, Revised and Expanded,edited by Seppo Salminen and Atte von
Wright
86. Handbook of Vegetable Science and Technology: Production, Composition,Storage,andProcessing,
editedby D. K.Salunkheand S. S.
Kadam
87. Polysaccharide Association Structures
in Food, edited by Reginald H.
Walter
88. Food Lipids: Chemistry, Nutrition, and Biotechnology, edited by Casimir
C. Akoh and DavidB. Min
89. Spice Science and Technology, Kenji Hirasa and Mitsuo Takemasa
90. Dairy Technology: Principles of Milk Properties and Processes,
P. Walstra, T. J. Geurts, A. Noomen, A. Jellema, and M. A.J. S. van Boekel
91. Coloring of Food, Drugs, and Cosmetics, Gisbert Otterstafter
92. Listeria, Listeriosis,andFoodSafety:SecondEdition,Revisedand
Expanded, edited by Elliot T. Ryser and Elmer
H. Marth
93. Complex Carbohydrates in Foods, edited by Susan Sungsoo Cho, Leon
Prosky, and Mark Dreher
94. Handbook of Food Preservation, edited by M. Shafiur Rahman
95. International Food Safety Handbook: Science, International Regulation,
and Control, edited by Kees van der Heijden, Maged Younes, Lawrence
Fishbein, and Sanford Miller
96.FattyAcids
in FoodsandTheirHealthImplications:SecondEdition,
Revised and Expanded,edited by Ching Kuang Chow
97. Seafood Enzymes: Utilization and Influence on Postharvest Seafood

Quality, edited by Norman F. Haard and BenjaminK. Simpson
98. Safe Handling of Foods, edited by Jeffrey M. Farber and Ewen C. D.
Todd
99. Handbookof Cereal Science andTechnology: Second Edition, Revised and Expanded, edited by Karel Kulp andJoseph G. Ponte, Jr.
100. Food Analysis byHPLC: Second Edition, Revised and Expanded,
edited by Leo M. L. Nollet
101. Surimi and Surimi Seafood, edited by Jae W. Park
102. Drug Residues in Foods: Pharmacology, Food Safety, and Analysis,
Nickos A. Botsoglou and Dimitrios J. Fletouris
103. Seafood and Freshwater Toxins: Pharmacology, Physiology, and
Detection, edited by Luis M. Botana
104. Handbook of Nutrition and Diet, Babasaheb B. Desai
105. Nondestructive Food Evaluation: Techniques to Analyze Properties
and Quality, edited by Sundaram Gunasekaran
106. Green Tea: Health Benefits and Applications, Yukihiko Hara
Additional Volumes in Preparation

Alternative Sweeteners: Third Edition, Revised and Expanded, edited
by Lyn OBrien Nabors
Handbook of Dietary Fiber, edited by Susan Sungsoo Cho and Mark
Dreher
Food Processing Operations Modeling: Design and Analysis, edited
by Joseph lrudayaraj
Handbook of Microwave Technology for Food Applications, edited by
Ashim K. Datta and R. C. Anantheswaran
Applied Dairy Microbiology: Second Edition, Revised and Expanded,
edited by Elmer H. Marthand James L. Steele
Food Additives: Second Edition, Revised and Expanded, edited by
John H. Thorngate 11, Seppo Salminen, and Alfred Larry Branen
Nondestructive
Food Evaluation
Techniques to Analyze Properties and Quality
edited by
Sundaram Gunasekaran
University of Wisconsin-Madison
Madison, Wisconsin
MARCEL
MARCEL
DEKKER,
1NC.
D E K K E R
N E WYORK BASEL
ISBN: 0-8247-0453-3
This book is printed on acid-free paper.

Headquarters
Marcel Dekker, Inc.

270 Madison Avenue, New York, NY 10016
tel: 2 12-696-9000; fax: 212-685-4540
Eastern Hemisphere Distribution
Marcel Dekker AG
Hutgasse 4, Postfach 812, CH-4001Basel, Switzerland
tel: 41-61-261-8482; fax: 41-61-261-8896
World Wide Web
http:Nwww.dekker.com
For more
The publisher offers discounts on this book when ordered in bulk quantities.
information,writetoSpecialSales/ProfessionalMarketingattheheadquartersaddress
above.
Copyright 0 2001 by Marcel Dekker, Inc. All Rights Reserved.
Neither this book nor any part may be reproduced or transmitted in any form or by any
means, electronic or mechanical, including photocopying, microfilming, and recording,
or by any information storage and retrieval system, without permission in writing from
the publisher.
Current printing (last digit):
I 0 9 8 7 6 5 4 3 2 1
PRINTED IN THE UNITED STATES OF AMERICA
Preface
In the food industry, we are inherently limited

by our inability to objectively,
consistently, and accurately test food quality by our faculties
of sight, sound,
of developtouch, taste, and smell. Fortunately, however, through many years
ment, we have sensors that assist, andin many cases replace, human evaluations.
Nonetheless, on-line control of food processes remains a major challenge in designing processes to consistently produce high-quality foods.
The recent development of new sensors and measuring techniques has created several new opportuin thisveryimportantaspectoffood
nities to assistthefoodindustry
manufacturing. Rapid, nondestructive, and on-line food quality evaluations can
improve plant productivity and cost-effectiveness. Therefore, it is a very critical
issue for the food industry. This book is a comprehensive treatise on most of the
nondestructive methods for food quality evaluation and
is designed to serve as
asinglereferencesourcefortheindustryandacademia.Emphasishasbeen
placed on the new and emerging methods and applications.
Nondestructive Food Evaluation is an edited volume with contributions
from active researchers and experts in their topic areas. The bookis divided into
10 chapters, each focusing on a major nondestructive techique, including optical,
magnetic, ultrasonic, mechanical, and biological methods. Each chapter informs
the reader of significant advances and offers insights for possible future trends
in the nondestructive method.
iii
iv
Preface
Optical techniques are presented under four topical headings (Chapters1of electromagnetic spectrum: visible, IR, NIR, and
FTIR; computer vision; delayed light and fluorescence; and x-ray tomography.
Chapter 5 introduces the basic principles of nuclear magnetic resonance (NMR)
and magnetic resonance imaging (MRI). NMR and MRI are nondestructive techniques that can be used to probe the physical and chemical properties and anatomical structure of biological materials. Therefore, the quality parameters associated
with certain physical and chemical properties of foods can be evaluatedby NMR
and MRI. Use of NMR and MRI in analysis of water mobility, glass transition
in foods is described.
process, distributionof water and fat, and internal blemishes
Sound waves are transmitted through materials as compressions and rarefactions in their physical structure. Hence, it is often possible to relate the ultrasonic properties of a material to useful information about its macroscopic and
microscopic composition. In Chapter 6, the physics of high-frequency sound is
introduced, and applications of ultrasonic properties to monitor food quality are
described.
Mechanical methods of nondestructive food evaluation include low-intensity impact (tapping) and vibration testing and high-pressure air impingement
(Chapters 7 and 8). One of the most recent techniques used to evaluate food
texture is the small-amplitude oscillatory strain test, popularly known
as dynamic
5%) is used to study the material
testing. In this test, a very small strain (less than
structure-function relationships. Since food structure is the basis for its texture,
this method offers the advantageof obtaining fundamental information about the
eating quality of foods.
A taste panel traditionally measures many subjective food quality factors
such as aroma andtaste.Recent developments in providing objective, instrumented evaluations of such subjective factors are presented in Chapter 9, Biosensors in Food Quality Evaluation. A good example of such class of sensor
is the electronic nose, which mimics the human sense
of smell. The integration
of multiple gas sensors and artificial intelligence has led to a
new science of
machine olfaction. Biosensors offer the advantage
of rapid detection of bioactive
componentsthatcanbemeasured
and controlledtoensurefoodquality
and
safety. In food quality analysis and control, the data collected often are subjective
and ill-conditioned. To infer useful information out
of such data sets requires
methods other than those traditionally used. Chapter 10 describes some of these
data analysis procedures, such as neural networks, fuzzy logic, pattern recognition, and statistical process control.
I would like to thankallthe contributors and the Marcel Dekker, Inc.,
production staff for their enthusiastic and timely support in bringing this project
to fruition.
4) to cover the wide span
Contents
Preface
Contributors
1.
Optical Methods: Visible, NIR, and FI'IR Spectroscopy

Sundaram Gunasekaran and Joseph Irudayaraj
2. Computer Vision
Suranjan Panigrahi and Sundaram Gunasekaran
3.
Delayed Light Emission and Fluorescence

Sundaram Gunasekaran and Suranjan Panigrahi
4.
X-Ray Imaging for Classifying Food Products Based on Internal

Defects
Ernest W. Tollner and Muhammad Afzal Shahin
5 . Nuclear Magnetic Resonance Techniques and Their Application

in Food Quality Analysis
R. Roger R u m and Paul L. Chen
...
111
vii
39
99
137
165
V
vi
Contents
6. Ultrasonics
John Coupland and David Julian McClernents
217
7. Firmness-MeasurementMethods
Yen-Con Hung, Stan Prussia, and Gabriel 0. I . Ezeike
243
8. LinearViscoelasticMethods
M . Mehrnet Ak and Sundararn Gunasekaran
287
9.Biosensors in FoodQualityEvaluation
Sudhir S. Deshpande
335
10. New Techniques for Food Quality Data Analysis and Control
Jinglu Tan
319
Index
417
Contributors
M. Mehmet Ak Department of Food Engineering, Istanbul Technical University, Maslak, Istanbul, Turkey
Paul L. Chen Department of Biosystems and Agricultural Engineering, University of Minnesota, St. Paul, Minnesota
John Coupland Department of Food Science, The Pennsylvania State University, University Park, Pennsylvania
Sudhir S. Deshpande Signature Bioscience, Inc., Burlingame, California
Gabriel 0.I. Ezeike Center for Food Safetyand Quality Enhancement, Department of Food ScienceandTechnology,TheUniversity
of Georgia,Griffin,
Georgia
SundaramGunasekaran
Department of BiologicalSystemsEngineering,
University of Wisconsin-Madison, Madison, Wisconsin
Yen-Con Hung Center for Food Safety and Quality Enhancement, Department
of Food Science and Technology, The University of Georgia, Griffin, Georgia
vii
viii
Contributors
Joseph Irudayaraj Department of Agricultural andBiologicalEngineering,

The Pennsylvania State University, University Park, Pennsylvania
David Julian McClements Department of Food Science, University of Massachusetts, Amherst, Massachusetts
Suranjan Panigrahi Department of Agriculture and Biosystems Engineering,
North Dakota State University, Fargo, North Dakota
Stan Prussia Department of Biological and Agricultural Engineering, The University of Georgia, Griffin, Georgia
R. Roger Ruan Department of Biosystems and Agricultural Engineering, University of Minnesota, St. Paul, Minnesota
Muhammad Afzal Shahin Grain Research Laboratory, Canadian Grain Commission, Winnipeg, Manitoba, Canada
Jinglu Tan Department of Biological and Agricultural Engineering, University
of Missouri, Columbia, Missouri
Ernest W. Tollner Department of BiologicalandAgriculturalEngineering,
Driftmier Engineering Center, The University of Georgia, Athens, Georgia
Optical Methods: Visible, NIR, and

FTlR Spectroscopy
Joseph lrudayaraj
1.
INTRODUCTION
Food quality may be defined as the compositeof those characteristics that differentiate individual units of a product and have significance
in determining the
degree of acceptability of that unit by the buyer (1). The qualityof many products
may be judged by the colors they display or fail to display. It is particularly
and vegetables,
critical in cases of food and biological products such as fruits
cereal grains, and processed foods. The primary goal
of quality control is to
maintain a consistent standard of quality at a reasonable cost and at levels and
tolerances acceptable to buyers.
Human evaluation has been the primary method of quality assessment for
operations such as grading and sorting of food products, but such evaluation can
hardly provide a general standardon a large scale and wide range of operations.
Factors such as eye fatigue, lackof color memory, variationsin color discriminain lightingconditions
tionbetweenindividuals,personalbias,andvariations
greatly influence an individuals decision when determining the qualityof a certain product. Moreover, the human eye is greatly limited by its perceptions in a
very narrow band of the vast electromagnetic spectrum. Some quality attributes,
external and internal defects, and compositional factors are more readily detect(UV) and infrared
able in the region outside the visible range, e.g., ultraviolet
(IR). This has led to considerable research in developing instruments sensitive
1
Gunasekaran and lrudayaraj
to a broad band of the electromagnetic spectrum and in establishing indices of

quality for various food and biological materials. The nondestructive nature
of
optical methods has made them particularly attractive in on-line quality evaluation involving a large number of samples for processing operations.
IR radiation that are
In this chapter optical methods based on visible and
routinely used in quality evaluation and control of several food products are discussed. A major emphasis isplaced on nondestructive quality evaluation methods
for agricultural and biological materials that are useful
in evaluating maturity
and/or ripeness, detecting external and internal defects, and composition analysis.
II. LIGHT ANDCOLORMEASUREMENT

Among the properties widely used for analytical evaluation
of materials, color
to possess a specific
is unique in several aspects. While every material can be said
property such as mass, no material is actually colored as such. Color is primarily
in a
an appearance property attributed to the spectral distribution of light and,
way, is related to some source of radiant energy (the illuminant), to the object
to which the color is ascribed, and
to the eye of the observer. Without light or
the illuminant, color does not exist. Therefore, several factors that influence the
(2):
radiation subsequently affect the exact color that an individual preceives
Spectral energy distribution of light
Conditions under which the color is viewed
Spectral characteristics of the object with respect to absorption, reflection,
and transmission
Sensitivity of the eye
Thus, in reality, color is in the eye of the observer, rather than in the colored object. The property of an object that gives it a characteristic color is its
light-absorptive capacity. Since lightis the basic stimulus of colors, it is important
to consider the electromagnetic spectrum (Fig. 1). Several optical methods have
been developed based on radiation from different regions of this spectrum.
Radiation is one of the basic physical processes by which energy is transferred from one place to another. Propagation of radiation through free space is
thesamefortheentireelectromagneticspectrum,
i.e., radiation of all wavelengths-from the shortest gamma rays to the longest radio waves-travels with
small partof the electromagthe same speed in vacuum. Visible light forms aonly
netic spectrum, with a spectral range from approximately 390 nm (violet) to 750
nm (red). The sensitivity of the eye varies even within this narrow visible range.
Under conditions of moderate-to-strong illumination, the eye is most sensitive
to yellow-green light of about 550 nm.
Optical Methods
Fig. 1 The electromagnetic spectrum.
If the spectral distribution throughout the visible region is unequal, then

the sensation of color is evokedby radiant energy reaching the eye's retina.An
equal spectral distribution makes the light appear as white. The unequal distribuof the source itself,
tion responsible for color sensation may be characteristic
such as flame spectra composed of one or more monochromatic wavelengths, or
may result from selective absorptionby the system, which appears colored. The
latter includes several systems that show selective absorption for light and exhibit
color as a resultof reflection or transmissionof unabsorbed incident radiant energy (Fig. 2). The three basic factors required in color sensation include the radiator or illuminant, the object, and the observer. The radiant energy emitted by the
radiator is characterized
by its spectral quality, angular distribution, and intensity.
Hutchings (3) lists the following material properties and lighting of the
scene as affecting the total appearanceof the object:
Material properties:
Optical properties (spectral, reflectance, transmission)
Physical form (shape, size, surface texture)
Temporal aspects (movement, gesture, rhythm)
Lighting of the scene:
Illumination type (primary, secondary, tertiary)
Spectral and intensity properties; directions and distributions
Color-rendering properties
lrudayaraj Gunasekaran and
Reflection
Specular
Radiation
Incident
Medium 1 , n,
n
Y
Medium 2, n2
Light Scattering
and Absorption
Diffuse kkmission\
\r
Regular Transmission
Fig. 2 Schematic representation of interaction of light with matter. 8, = angle of incidence, eR = angle of reflectance, OT = angle of transmittance, n,, n2 = refractive index
of medium 1 and 2, respectively.
A.
Color Specification
There are three characteristics of light by which a color may be specified: hue,
saturation, and brightness. Hueis an attribute associatedwith the dominant wavelength in a mixture of light waves, i.e., it represents the dominant color as perceived by an observer. Saturation refers to relative purity or the amount
of white
light mixed with a hue. Brightness is a subjective term, which embodies the
chromatic notion of intensity. Hue and saturation taken together are called chromaticity. Therefore, a colormay be characterized by brightness and chromaticity
(4).
B. CIE System
The Commission de Internationale de 1Eclairage (CIE) defined a system
of describing the color of an object based on three primary stimuli: red
(700 nm),
green (546.1 nm), and blue (435.8 nm). Because of the structure of the human
eye, all colors appear as different combinations of these. The amounts of red,
form any given color are called the tristimulus
green, and blue needed to
values, X, Y, and Z, respectively. Using theX, Y, and Z values, a color is represented by a set of chromaticity coordinates or trichromatic coefficients,x, y, and
z, as defined below:
Optical
x =
X
X + Y + Z
Y
y = x + Y + z
z =
X + Y + Z
(1)
It is obvious from the equations above thatx y + z = 1. The tristimulus values

for any wavelength can be obtained from either standard tables or figures.
A plot
that representsall colors in x (red)-y (green) coordinatesis known as a chromaticity diagram. For a givenset of x and y, z is calculated from the above equations.
Therefore, colors are generally specified in terms of Y, x, and y.
There are a number of color metrics based
on the CIE system. They include
CIE Lightness, CIELUV, CIELAB, etc.
In the food industry, the CIELAB system
has been popular. For example, objective measurements of color using the CIELAB color parameters such as L* (lightness), a* (redness), and hue angle have
been used to evaluate pork quality on-line in a industrial context (5,6).
Other color models, suchas the RGB, CMY, and HSI, etc., are
very similar
to the CIE system, and numerical representation of a color in one system can be
converted into another (4).
C. MunsellSystem and Atlas

The Munsell color-order system isway
a of precisely specifying colorsand showing the relationships among colors. Every color has three qualities or attributes:
hue, value, and chroma. A set of numerical scales with visually uniform steps
for each of these attributes has been established. The Munsell Book of Color
displays a collection of colored chips arranged according to these scales. Each
chip is identified numerically using these scales. The color
of any surface can
be identified by comparing it to the chips under proper illumination and viewing
conditions. The color is
then identified by its hue, value, and chroma. These
attributes are given the symbols H, V, and C and are written in a form H V/C,
which is called the Munsell notation. Using Munsell notations, each color has a
logical relationship toall other colors. This opens up endless creative possibilities
in color choices, as well as the ability to communicate those color choices prein food
cisely. The Munsell systemis the color order system most widely quoted
industry literature (3). Food products for which the U.S. Department of Agriculture (USDA) recommends matching Munsell discs to used
be include dairy products such as milk and cheese, egg yolks, beef, several fruits, vegetables, and fruit
juices (3).
Other color atlases and charts are available for use in the food industry,
such as the Natural Color System and Atlas, Royal Horticultural Society Charts,
etc. (3). These atlases and charts are used for visual comparison of a product
color with that of a standard color (diagram), which is still commonly practiced
in the food industry. The evaluationof potato chip color is a very good example.
111.
INTERACTION OF LIGHTWITHMATTER
A.
PhysicalLaws
When light falls on an object, it may be reflected, transmitted, or absorbed (Fig.

2 ) . Reflected light is the partof the incident energy that is bouncedoff the object
surface, transmitted light passes through the object, and absorbed light constitutes
the part of the incident radiant energy absorbed within the material. The degree
to which these phenomena take place depends on the nature of the material and
on the particular wavelength of the electromagnetic spectrum being used. Commonly, optical properties of a material can be defined by the relative magnitudes
of reflected, transmitted, and absorbed energy at each wavelength. Conservation
of energy requires that sum of the reflected (IR), transmitted (IT), and absorbed
(IA) radiation equals the total incident radiation (I). Thus,
IR
+ IT + IA
(2)
According to its transmittance properties, an object may be transparent,

opaque, or translucent. Almost all food and biological products may be considered to be opaque, although most transmit light to some extent at certain wavelengths (7). The direction of a transmitted ray after meeting a plane interface
between any two nonabsorbing media can be predicted based on Snells law:
(3)
n2 sin 8, = n , sin 8,
The attenuation of the transmitted rayin a homogeneous, nondiffusing, absorbing

medium is defined by Beer-Lamberts law:
where k is a constant and n is the number of molecules in the path of the beam.
of the sample and the thickness b
Since n is proportionaltoconcentrationc
through which the radiation passes, Eq. (4) is rewritten as:
(5)
lOg(IT/I) = abc
The ratio IT/Iis known as the transmittance T and is related to absorbance A as:
A = log( 1/T)
From Eqs. (5) and (6), absorbance A can also be written
A = abc
as:
(7)
where a is called the absorptivity. [If c is expressed in mol/L and b in cm, a is

replaced by the molar absorptivity, E (L/mol . cm).]
Various constituents of food products can absorb a certain amount of this
radiation. Absorption varies with the constituents, wavelength, and path length
of the light (8). The absorbed energy can be transformed into other forms of
Optical Methods
energy such as fluorescence, phosphorescence, delayed light emission, heat, etc.

Optical density (OD) is more commonly
used to describe absorption. OD has
the same definition as that of A, i.e., OD = log( l/T)],but it is used for applications where the transmitted ray is attenuated by both geometrical means (scatter)
and absorption. The advantages of using the OD scale are as follows (9). First,
the analysis is simpler, i.e., an optical density difference, AOD, is equivalent to
a transmittance ratio. The differences are easier
to compute. (Note: AOD(A B) = log( 1/RA) - log( 1/RH) whereA and B are wavelengths at which the measurements are made.) Second, the logarithmic plot permits a wider of
range
intensities, i.e., a transmittance scale of 1 to 100 is two orders of magnitude, while
an OD scalemay cover five orders of magnitude. Finally, there aislinear relationship between OD and the concentration of an absorbing substance.
Reflection is a complex action involving several physical phenomena. Depending on how light is reflected back after striking an object, reflection may be
2). Reflection
defined as regular or specular reflection and diffused reflection (Fig.
from a smooth, polished surface
is called specular or regular. It mainly
produces the gloss or shine of the material (1,2,10,1I). The basic law of specular
reflection states that the angle at which a ray is incident to a surface must equal
the angle at which it is reflected off the surface. Fresnel equations define the
phenomenon of specular reflection. The intensityof parallel Ril and perpendicular
RI components of the reflected light are:
RII
(n2/nl)?cos 8, - [(nz/nl)- sin? 8J1/?

(n2/nI)>cos 8, + [(nz/nl)?- sin2 81]1/2
cos 8, - [(nz/nl)?- sin? 8,]/?
cos el + [(n?/n,)?- sin2
The regular reflectance R = Ri

RI, and hence
+ R:
and for normal incidence (8 =
OO),
Rii =
where n , and n z are refractive index of the medium and object, respectively; and
8, is the incidence angle (Fig. 2).If the material is absorbing, the refractive index
is a complex number n( 1 - ik), where n is the real part of the complex number
and k is an absorption constant, and the regular reflectance is written as:
=
[I.?
(n?
nJ2+ ( n M ]
+ n,)? + (nzk)2
(1 1)
When the incident light is reflected from a surface evenly at all angles, the
object appears to have a flat or dull finish termed diffuse reflection. No rigor-
and
Gunasekaran
lrudayaraj
ous theory has been developed for diffuse reflectance, but several phenomenological theories have been proposed, the most popular being the Kubelka-Munk theory (12). The Kubelka-Munk model relates sample concentration to the intensity
way Beer-Lamberts
of the measured spectrum in a manner analogous to the
law relates band intensities to concentration for transmission measurements. The
Kubelka-Munk function f(R,) is generally expressed as:
f(RJ =
(1 - R-)2 - k
S
2R,
where R, = absolutereflectance of an infinitelythicklayer,k

= absorption
coefficient, and s = scattering coefficient.
Kubelka-Munk theory predicts a linear relationship between spectral data
and sample concentration under conditionsof constant scattering coefficient and
infinite sample dilution in a nonabsorbing matrix such as KBr (potassium bromide). Hence, the relationship can only be applied to highly diluted samples in
a nonabsorbing matrix. In addition, the scattering coefficient is a function
of
particle size, so samples must be prepared to a uniform fine size if quantitatively
valid measurements are desired.
It is not easy to quantify diffuse reflectance measurements since sample
transmission, scattering, absorption, and reflection all contribute to the overall
effect. By reducing particle size and dilution in appropriate matrices, surface
reflection that can give strong inverted bands is reduced and the spectra more
closely resemble transmission measurements. Typically, quantitative diffuse rein log( 1/R) units, analogous to absorbance
flectance measurements are presented
log( 1/T) units for transmission measurements. Bands increase logarithmically
with changes in the reflectance values.By comparison, bandsin spectra displayed
of reflectance. This differin Kubelka-Munk unitsvary as a function of the square
ence emphasizes strong absorbance bands relative to weaker bands.
The diffuse reflectance may be measured with respect to a nonabsorbing
standard and converted to produce a nearly linear relationship
with concentration
c as follows (13):
log(R/R) = log(l/R)
+ log(R) = a d s
(13)
where R and R = reflectance of the standard and the sample

(R > R), a =
absorptivity, c = concentration, and s = scattering coefficient. For monochromatic radiation, log R is constant and may be ignored, and Eq. (13)
may be
written as (12):
(14) log(l/R)
(s/a)
c =k
where k = absorption coefficient. It should be noted that s is not a constant but
depends on a number of properties of the sample such as particle size (s is in-
Optical Methods
versely proportional to particle size) and moisture content. Equation (14)

is the
basis for near-infrared(NIR) spectroscopic analysisof foods (14).In food materials, the primary factor that influences light reflection is a phenomenon known
as
scattering or diffusion (2,7,10).If the surface of incidence is rough, incident light
will be scattered in all directions. Since the incident rays strike a rough surface
more than once before being reflected, they would be expected to have a lower
total reflectance than those reflected from a smooth surface (15).
In classical optics, diffuse reflection was thought
be to
responsible for color.
It was also commonly believed that colors of natural objects, such as foods and
plant foliage, are seen by means of light reflected off their surfaces. Birth (15)
recognized that the light must be transmitted through pigment within the cells
in order to produce a colored appearance. Since
most food materials are optically
in all directions. Only
nonhomogeneous, light entering such material is scattered
about 4-5% of the incident radiationis reflected off the surfaceof these materials
as regular reflectance (7,16). The remaining radiation transmits through the surface and encounters small interfaces in the cellular structure and reflects back.
A large fraction of this reflected radiation from within the material is scattered
back to the surface through the initial interface. This type
of reflection is termed
as body reflectance (7). The body reflectance is nearly always diffuse and is
the most significant form of reflectance for foods. Some part of the transmitted
light diffuses deeper into the material
and may eventually reach the surface some
distance away from the incident point.
B. FactorsAffectingDiffuseReflectanceSpectralData
Diffuse reflectance spectroscopy offers exceptional versatilityin sample analysis.
This versatility results from both its sensitivity and optical characteristics. Classically, diffuse reflectance has been used to analyze powered solids in a nonabsorbing matrix of an alkali halide such as KBr. The sample is typically analyzed
at low concentrations, permitting quantitative presentation
of the datain KubelkaMunk units. This technique yields spectra that are qualitatively similar to those
produced by conventional transmittance or pellet methods. However,
they exhibit
higher sensitivity for quantificationand are less subject to scattering effects that
cause sloping baselines in pellet measurements.
Several factors determine band shape and relative/absolute intensity
in diffuse reflectance spectroscopy through their effect on the reflection/absorbance
phenomena specific to the sample. These include:
Refractive index of the sample
Particle size
Sample homogeneity
Concentration
10
1.
RefractiveIndex
Refractive index affects the results via specular reflectance contributions to diffuse reflectance spectra. With organic samples, the spectra display pronounced
changes in band shape and relative peak intensities, resulting in nonlinearity in
For some inorthe relationship between band intensity and sample concentration.
ganic samples, strong specular reflection contributions can even result in complete band inversions. This overlayof diffuse reflectance and specular reflectance
by diluting
spectra, as well as the resulting spectral distortions, can be minimized
the sample in a nonabsorbing matrix. In addition, accessory design can help reduce specular reflectance contributions.
2. ParticleSize
Particle size is a major consideration when performing diffuse reflectance measurements of solids. The bandwidth is decreased and relative intensities are dramatically altered as particle size decreases. These effects are even more pronounced in spectra of highly absorbing inorganic materials with high refractive
indices. For these samples, specular contributions can dominate thefinal spectra
if the particle size is too large. To acquire a true diffuse reflectance spectrum, it
is necessary to uniformly grind the sample and dilute it in a fine, nonabsorbing
matrix. Similar preparation must be applied to the nonabsorbing matrix material
in order to provide and ideal diffuse reflector for background analysis and as
a support matrix for the sample.
3. SampleHomogeneity
The Kubelka-Munk model for diffuse reflectance is derived for a homogeneous
sample of infinite thickness. However, some sample analysis methods, especially
of sample onto a powdered
those designed for liquid samples (e.g., deposition
supporting matrix), can result in a higher concentration of sample near the analysis surface. In these circumstances, variations in relative peak intensities may be
noticed. In particular, more weakly absorbing wavelengths tend to be attenuated
it is
at higher sample concentrations. To avoid these peak intensity variations,
necessary to distribute the analyte as uniformly as possible within the nonabsorbing background matrix.
4.
Concentration
One particularly important advantage of diffuse reflectance spectroscopy, especially in comparison to transmittance measures, is its extremely broad sampleanalyzing range. While it is theoretically possible to acquire usable diffusereflectance spectraon samples of wide-ranging concentrations, practical considerations
often complicate the analysis process. With high concentration samples, espe-
Optical Methods
11
cially those with a high refractive index, one can expect a dramatic increase in
the specular contribution to the spectral data. As a result, some sample data may
be uninterpretable without adequate sample dilution. Even when samples can be
measured satisfactorily at high concentrations, it is advisable to grind the sample
to a very uniform and
fine particle size to minimize both specular reflectance
and sample scattering effects, which adversely affect quantitative precision.
Alternative methods of sample analysis in diffuse reflectance include:
of a solid in the presence of a nonabEvaporation of volatile solutions

sorbing supporting matrix
Deposition of a liquid sampleor dissolved solid onto the surface of a nonabsorbing supporting matrix asin analysis of liquid chromatography eluent
Direct analysis of certain solid samples, which has been successfully employed on a broad array of sample types including starch, wool cloth,
paper, plant leaves, pharmaceutical tablets, and cedar wood siding
IV.
NIRAND FTlR SPECTROSCOPY
A.
Near-InfraredSpectroscopy
IR spectroscopy has been used as an analytical technique for almost a century

(17). The IR region of the spectrum spans 0.780-1000 ym and has been divided
into near-, mid-, and far-IR subregions.
The most widely used are NIR, which
is from about 1 to 2.5 ym, and mid-IR, which is from 2.5 to 14.9 pm (18). In
IR studies, the frequency is often expressed in wave number (cm"), which is
the inverse of wavelength when expressed in centimeters.
IR spectroscopy is a form of vibrational spectroscopy but arises from an
interaction of IR radiation with molecular bonds within a sample (19). Any sample will absorb at a certain wavelength, depending on the characteristics of the
chemical entities present.The IR spectrum would reveal the particular absorption
band(s), which can be related
to the constituents present. It can be shown that
the IR radiation of frequency and energy
hu can supply energy required for a
transition provided that:
where 2) = frequency, h = Planck's constant, k = force constant, p = reduced

mass = (m, + m,)/(m, + m?), and m , and mz = mass of two atoms joined by
the bond being studied.
The above equation shows that the absorption frequency for a given bond
For
depends upon its strength and the masses of the atoms forming the bond.
example, though the bonds
C - 0 and C = O haveidenticalreducedmasses,
12
C -0 absorbs at a different frequency thanC=O. The C =O bond has a higher

force constant and a higher absorption frequency. The C-H
bond has a much
lower reduced mass and absorbs even at high frequencies. IR spectroscopy can
thus be used to determine which functional groups are present in a sample. Different functional groups in foods absorb IR radiation at different wavelengths
(Table1).
The constituents also affect the overall spectrum since scattering depends
on the ratio of the refractive index of the material n , to that of the surrounding
medium n2. For example, as the moisture content increases, so does the partial
pressure of water vapor around the particles. Since the refractive index of water
is greater than that of air, it leads to a decrease in n,/nz. Hences, the scattering
of these procoefficient, and therefore log(1/R) increase (14). The overall effect
cesses is that s becomes anew unknown for each sample. Therefore, the analytical
utilization of diffuse reflectance spectra must be carried out an
onempirical basis.
Also, since s varies in a complex mannerwith wavelength, background correction
in the case of diffuse reflectance spectra is difficult, although the basic principles
are the same as those outlined earlier. The artof NIR spectroscopy of scattering
samples lies in selecting measurement and reference wavelengths at which s is
nearly equal, so that the s/a term in Eq. (14) becomes a constant. Osborne and
Fearn (12) presented the theory of NIR spectrophotometry in much detail.
Commercial NIR instruments are manufactured using oneof three geometries to collect reflected light from samples: integrating sphere, large solid angle
detector, and small detector. The large solid angle detector offers good collection
efficiency, simplicity of construction, and minimum interference from specular
reflectance.
Table 1 CharacteristicFunctionalGroups ofFoodComponents
component
Chemical
Food
functionality
Wavelength
Water, carbohydrates
Unsaturated fat
Fats, proteins, carbohydrates
Fats
Pectin
Fatty acids, acetic acid
Water
Protein
Protein
Carbohydrates, fats
Source: Adapted from Ref. 19
of band
0 -H stretch
C -H of cis double bond
C-H
C=O, ester
C=O, ester
C =0, acidic
0 - H (bend)
C =0, amide I
N -H, amide I1
c-0, c-c
3600-3200
3030
3000-2700
1745
1725
1600- 1700
1640
1650
1550
1400-900 (complex)
Optical
13
B. FourierTransformInfraredSpectroscopy
Fourier transform infrared (FTIR) spectroscopy is based on the Michelson interferometer configuration designed a century ago (1 8,20,2I). It is used to detect
radiation in the mid-IR region. Fourier transform instruments obtain the data by
using interferometry while they calculate the spectrum by using Fourier transform
mathematics. The resultis increasing sensitivity of measurement.The interferometer consists of a fixed mirror, a movable mirror, and a beamsplitter (Fig. 3). The
beamsplitter transmits half the incident IR radiation to the movable mirror and
reflects the other half to the fixed mirror. The speed of the movable mirror is
monitored by a laser. The two mirrors reflect the two light beams back to the
beamsplitter and then the beams recombine. When the distance from the fixed
to the beammirror to the beamsplitter equals that from the movable mirror back
splitter, the amplitudes of all frequencies are in phase and recombine construca condition called zero
tively. There is no difference between the two beams,
retardation. As the movable mirror is moved away from the beamsplitter (retarded), the difference between the two beams
is generated because the two beams
travel different distances within the interferometer before recombining.A pattern
of constructive and destructive interference results, which depends
on the position
of the movable mirror and the frequency of the retardation. The intensity of the
Y
Source
Movable Mirror
Fig. 3 Schematicdiagram of theMichelsoninterferometer.
14
radiation is altered in a complex pattern as a functionof mirror movement. Thus,

the output radiation is modulated by the interferometer. Such recombined and
modulated IR radiation is directed through the sample compartment to the detector. It will generate a continuous electrical signal called an interferogram at the
detector.Acomputerisused
to changetheinterferogramintoasingle-beam
spectrum by a Fourier transform. In FTIR spectroscopy, the (background) spectrum of the source is first collected and stored at the computer. Then the sample
is placed in the sample compartment, and the spectrum is collected proportional
to the background spectrum to obtain the desired transmission spectrum.
FTIR is a rapid, precise qualitative technique for identifying and verifying
chemical compounds in foods with spectral multiplexing and optical throughput
(22,23). The procedure to prepare samples is not as complicated as that of traditional wet chemical procedures (24). With these advantages, FTIR can also be
used to determine food and food ingredient authentication (25). Powerful modern
data processing techniques, especially multivariate analysis, have been applied
to extract useful information from spectral data. However, water can be a problem
in F U R measurements because it absorbs strongly in regions about 3300 cm"'
of other chemical groups. Thus,
and 1600 cm", which overlap the absorption
some sampling procedures and computer software tools must be applied to overcome such problems.
1. AttenuatedTotalReflectionSpectroscopy
Attenuated total reflection (ATR) spectroscopy is one of the most powerful FTIR
methods for biological and liquid sample analysis. It is fast and yields a strong
signal even with small traces of the target molecule.
Reflection occurs when a beam of light passes from a dense to a less dense
medium. Total reflection occurs when the incident angle is greater than a critical
angle (Fig. 4). A light beam actually penetrates a small distance into the less
dense medium before reflection occurs(18). The depthof penetration varies from
a fraction of a wavelength up to several wavelengths. The depth of penetration
depends on the wavelength of the incident radiation, the index of refraction of
the two media, and the angle of the incident beam with respect to the interface.
Such penetration radiation is called the evanescent wave. When the less dense
medium absorbs the evanescent radiation, attenuation of a beam occurs at different wavelengths of the absorption bands. This is referred to as attenuated total
reflectance.
In the ATR method, the sampleis placed in contact against a special optical
crystal, which is called an internal reflectance element. An IR beam from the
spectrometer focused onto the beveled edgeof a setof mirrors is reflected through
the crystal, usually numerous times, and then is directed to the detector. Penetration d, is calculated as (26):
15
Optical Methods
"2
Fig. 4 Illustration of attenuated total reflectance. 8, = angle of incidence. 8,

angle, n , , 11: = refractive index of crystal and surrounding, respectively.
critical
where h = wavelength of the radiation in the internal reflectance element, 0 =

angle of incidence, n,p = ratio of the refractive indices of the sample vs. internal
reflectance element, and n p = refractive index of the internal reflectance element.
For a typical ATR setup, d, is
in the range between 10 and 20% of the
wavelength used (26). For quantitative purposes, FTIR-ATR can only be used
for homogeneous samples (27). Thus, the penetration depth
is typically 0.1-5
pm (28). When the incident angle
is changed, the penetration depth can also
be changed. Factors affecting the determination in an ATR experiment include
wavelength of the IR radiation, refractive index of the crystal and sample, depth
of penetration, effective path length, 670-4000 cm" angle
of incidence, efficiency of sample contact, and ATR crystal material (29).
2. Diffuse Reflectance Infrared Fourier Transform
Spectroscopy
Diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) is a sampling technique developed forIR analysis of powder materials and turbid liquids.
This technique has been applied for the analysisof pharmaceutical (30) and food
(3 1 ) samples. DRIFTS is also utilized for quantitative (32) and qualitative(33,34)
determination of the composition of samples (forages and coffee, respectively)
with an accuracy equal to or better than that found using NIR spectra.
DRIFTS has been used primarily in UV-visible spectroscopy studies where
the beam energy is high enough for proper detection. Diffuse reflectance studies
found little use in the caseof classical scanned IR beam. The advantages brought
16

Two elliptical
mirrors
IR beam
source
Optical
detector
Fig. 5 Schematic of DRIFTS with diffuse reflectance accessory.
by introducing Fourier transform methods (sensitivity and good signal-to-noise

to IR studies as
ratio) allowed the application of diffuse reflectance technique
well. The most commonly used device for collecting diffusely
reflected UV-visible and even NIR radiation froma sample is an integrating sphere whose interior
is coated witha diffusing, nonabsorbing powder suchas MgO (magnesium oxide)
or BaSOJ (barium sulfate) (Fig. 5). The sample and detector are usually held at
the surface of the sphere, and the measured spectrum is independent, to a good
approximation, of the spatial distribution of the reflected light and the relative
position of the sample and the detector.
DRIFTS offers several advantagesas a sample analysis technique, such as:
Minimal or no sample preparation
Very high sensitivity
Applicability across a wide range of sample concentrations
Ability to analyze most nonreflective materials including highly opaque or
weakly absorbing materials
Ability to analyze irregular surfaces or coatings such as polymer coatings
or glass fibers; suitability for very large, intractable samples through the
use of specialized sampling devices
V.
QUALITYEVALUATIONOFFOODPRODUCTS
A.
MaturityandRipenessEvaluation
The ripening of fruits is associated with changes in color, flavor, and texture that
lead to the state at which the fruit is acceptable to eat. Readily apparent phenom-
Optical
17
ena associated with the ripening of most fruits, among others, include changesin
color, which involve loss of chlorophyll leading to the unmaskingof underlying
pigments and the synthesis of new pigments. Ripening is regarded as an indication of senescence accompanied by several physiological and chemical changes
(12,13). It may also represent a process requiring synthesisof specific enzymes.
Most fruits exhibit an increased reflectance and reduced absorbance in the 670
nm region because of the loss of chlorophyll. This hasbeenthe single most
(35). The mechanismof color
important criterionin optically judging fruit quality
in Hutchings
changes in fruits and vegetables during ripening is further discussed
(3).
Determining degree of maturity by surface color evaluation, however, has
its limitations. In many fruits, external color changes donot reflect internal ripeness. Similarly, external color evaluation cannot truly differentiate between immature and mature green fruits. Such a distinction
is important because mature
green fruits eventually ripen while immature fruits
will not. Where skin color
does not truly represent fruit quality or does not sufficiently indicate the stage
of maturation, internal flesh color can serve as an index of quality (36).
The success of establishing a valid index of quality largely depends on
the appropriateness of the optical property measured. Basic requirements for a
successful optical measurement are that the magnitude of relative measurement
should vary as greatly as possible over the full range
of maturation, andthe
change in the measurement between consecutive stages of maturity should be
great enough to permit precise color differentiation.
In addition, the nature of
the measurement criterion should also be formulated judiciously so as to permit
measurements insensitive to variations in the product and the measurement system. Over the years, the measurement criterion has taken various forms, which
can be broadly grouped under the following four classes:
1.
Single wavelength measurement-the optical property

of the object at
a particular wavelength is considered as an index.
2. Difference measurement-the change in an optical property of the object at two wavelengths is measured. This is an indication of the average slope of the curves in the region between the two wavelengths.
The region should be chosen so as to obtain the maximum possible
change. The effectof variations in the object and apparatus are largely
attenuated in this kind of measurement (37).
3. Ratio measurement-the ratio
of an optical property at two or more
wavelengths is the chosen criterion. Generally, this is independent of
the sensitivitity of the measuring instrument.
4. Combination measurement-any combination of the above categories
is used. For instance, to evaluate the maturityof tomatoes, Birth et al.
(38) suggested the form
18
To identify diseased potatoes, Muir et al. (39) used
where the subscripts A, B, C, and D represent wavelengths at which

the transmittance T and reflectance R measurements are made.
In apples, major changes in spectral characteristics were found to occurin
the visible region. As the fruits mature, the percentage of reflectance increases
at about 676nm and greatly decreases between 400 and 600
nm (35). Peak reflectance was observed at 800 nm, but it decreased thereafter, presumably due to
absorption by water in the NIR region (35). Since the water content in the skin
is not likely to change with maturity, the reflectance would not change in the
NIR range. A ratio measurement, RSKO/Rh2",was found to be the best for red
varieties of apples. The Golden Delicious variety can be evaluated for maturity
at a single wavelength in the 550-620 nm region. Yeatman and Norris (40) suggested an OD difference (AOD) between 740 and 695 nm for several varieties.
As in apples, themost noticeable maturity-induced changein peaches is the
band at 675 nm, indicating a
decrease in absorbance in a fairly narrow wavelength
decrease in chlorophyll content (41). A decrease in absorbance at 675 nm was
also accompanied by a smaller decrease over a wider wavelength area beginning
with
at about 550 nm. This latter decrease gradually shifted to larger wavelengths
increasing ripeness of peaches.
Although measurements suchas those described above are useful to follow
maturation and ripening within varieties, they are less suitable as a general maturity index for the whole class
of fruits. Color differences among varieties
of
in wavelength of peak transmittance,
peachesarelikelytocausedifferences
which might not specifically be related to the
stage of maturation. The optical
density differences between two wavelength readings may be better suited for
in
such purposes. As mentioned earlier, such differences overcome differences
fruit size and variations induced by instrumental factors. Sidwell et al. (41) studied three such AOD values-AOD (700-720), AOD (700-740), and AOD (700750)-for peaches harvested at different stages of maturity. Such measurements
were apparently stable and were not influenced by the chlorophyll content. The
maturity index AOD (700-740) yielded the highest correlation ( r = 0.957) with
eating quality peaches. (Note: The numbers
in parenthesis following AOD and
subscripts used with T and R refer to wavelength in nanometers. This notation
continues in the rest of the text.)
Optical
19
The CIE tristimulus values X, Y, and Z or the chromaticity coordinates x,

of objects, can also be
y, and z, which are normally used to express the color
used to fonnulate general maturity indices for various fruits. This is particularly
true if the reflectance variation related to the stage
of maturation occurs in the
(35) studiedtherelationshipbetweenthese
visibleregion.BittnerandNorris
parameters and the picking date for several varieties
of apples, peaches, and pears.
ReflectanceratiosRSxo/RhzoandR670/R730seemedpromising as indicatingthe
stage of maturation for most of fruit varieties investigated.
For fruits like tomatoes, which are processed into several forms (e.g.. puree,
juice, or paste), internal color is far more important than skin color. Birth et al.
(38) reported a nondestructive measurement of internal colorof tomatoes by speca minimum transmittance at 670
tral transmittance. Green tomatoes exhibited
nm, which increased by more than five decades as they matured to the red stage,
corresponding to loss of chlorophyll. On the other hand, an equivalent relative
decrease in transmittance at 550 nm was observed, corresponding to the increase
in lycopene, the characteristic red pigment of tomatoes. The measurement Th2,)/
Th7[)
was found to change on the order of 30: 1 as tomatoes ripened from yellow
to full red color. These two wavelengths (620 and 670 nm) were specifically
selected because the spectral transmittance change was the greatest at 620 nm
as the fruit developed a redder color.
At 670 nm, the transmittance value was
high enough that an extremely sensitive instrument was not required. Similarly,
T520/TSJs value could distinguish externally green tomatoes with internal amber
color from tomatoesthat were green throughout. Combining the above two criteto indicate
ria, Birth et al. (38) proposed a new ratio, (T670 - Ts2~J/(Thz,)TSJS),
a very high correlation ( r =
internal color of tomatoes. This new index gave
0.95) with the color of the extracted juice.
The Agtron Grade G given by Eq. (1 6) is used by the food industry in
California to grade raw tomatoes based on reflectance measurements made using
the Agtron colorimeter (2):
Maturity and other quality indices for several other fruits and vegetables have
been developed following a similar pattern. A detailed list is available in Gunasekaran et al. (42).
IR and NIR methods can beused to determine the sugarsin fruits that have
a thin skin, such as apples, peaches, prunes, or cherries, which, in turn, can be
correlated to their stateof ripeness (43). Sensors basedon this principle have been
developed. However, this method is ineffective for thick-skinned fruits (44,45).
Optical properties have been used to evaluate maturity in peanuts. Kramer
et al. (46) investigated light absorption properties
of Virginia-type cured, un-
20
roastedpeanuthalveswithoutskins.AOD(480-510)wasfoundusefulasan
indicator of peanut maturity. Due to the absorption band at 445 and 470 nm of
the pigment xanthophyll in immature peanuts, the above measurement values
also found
werehigherforimmaturepeanutsthanformaturepeanuts.They
the tastepanelratingof
offgoodcorrelationbetweenAOD(480-587)and
flavor in peanuts. Beasley and Dickens (47) indicated that
oil extracted from
as amaturityindex.Theyreportedthatoilfrom
peanutscouldalsobeused
mature peanuts generally transmitted more light at about 425, 456, and 480 nm
than oil from immature peanuts. Apart from xanthophyll, p-carotene and lubein
to
were also suggested as influencing the spectral properties, which are related
maturity.
Gloss characteristics of a number of fruits and vegetables have been determined. Unwaxed oranges, bananas, and onions have significantly lower gloss
than eggplant, green pepper,and tomato (48). Commercially mature eggplants are
glossier than green tomatoes and apples. This
is partly explained by differencesin
epicuticular wax structure. The lamellae-type wax covering the eggplant reflects
light more efficiently than the amorphous wax layer covering the tomato and the
as they
large overlapping platelets of the apple (49). Bananas also lose gloss
mature. This is probably due to epicuticular wax from the surface and possibly
also due to an increase in surface roughness caused by water transfer by osmosis
to shriveling, separationof epidermal cells,
from the skin to the pulp, which leads
and the appearance of longitudinal cracks (50).
6. Detection of External and Internal Defects
7. Fruits and Vegetable Products

Optical properties of fruits and vegetables are affected by external and internal
defects, including mechanical injuries that occur during harvesting and postharvest handling as well as certain microbial diseases. Some common problems encountered in mechanical fruit harvesting include skin damageor bruising, as well
as association of nonedible, unwanted parts such as stem and calyx. Tender fruit
varieties that make up a very large portion
of fresh market fruit production are
especially susceptible to bruising during mechanical harvesting. Fruit color and
bruise appearance may change substantially between the time
of harvest and final
grading. By measuring light reflected from such defects and comparing it with
that reflected from the undamaged surface, defective fruits can be sortedout from
normal fruits. Similarly, measuring transmittance characteristics of fruits could
help identify several internal defects.
Bruising on apples is a major problem in grading operations. With the increasing use of mechanical hurvesters, the numbers of bruises and other surface
defects are expected to increase. Attempts to develop an automatic apple bruise
Optical Methods
21
detection device to sort out bruised fruit beganin the early 1970s. While studying
the rate of discoloration for impact injuries versus the changesin selected phenolic compoundsin bruised apple pulp, Ingle and Hyde( 5 l ) observed a consistently
(52j reported a
lower reflectance at 600 nm for bruised apple pulp. Woolley
decrease in the NIR diffuse reflectance when water was used to replace air in the
intercellular air spacesin plant tissues. Since an apple bruise primarily consists
of
crushed cells surroundedby free liquid, measuring NIR reflectance seems promis(53)
ing to detect bruising in apples. Using a similar technique, Brown et al.
extensivelystudiedtheNIRreflectance
of threevarieties of freshandstored
apples. At all wavelengths between 700 and 2200 nm, bruised apple skin exhibited a less average reflectance than unbruised skin. It is thought that reflectance
a combifor a bruised surface is less than that for an unbruised surface because
of
nation of cell destruction in the bruise (fewer rigid cell walls to scatter light), an
in
unaltered water-air relationship in the tissue, and a gradual chemical change
the cell material. The differences
in reflectance at 800, 1200, and 1700 nm or
the ratio of reflectance (unbruised to bruised) at some wavelength between 1400
and 2000 nm may be useful in optical bruise detection (53). Reflectance properties of apple tissue cannot reliably predict bruise depth, but two-wavelength derivative models distinguish between good and bruised tissue better than nonderiva( 5 5 ) alsoinvestigatedtypicalreflectanceproperties
of
tivemodels(54).Reid
three varieties of apples for use in automatic trimming operations. He determined
that a detector sensitive in the wavelength range from 400 to 450 nm could be
used for bruise detection, and one sensitive in the 725-800 nm region could be
used for stem and calyx detection.
in
Reflectance properties have also been used to detect water core defect
apples. a heat-initiated disorder. In affected tissue, intercellular spaces are filled
with liquid or cells become swollen so as to eliminate air spaces, resulting in a
translucent or water-soaked appearance. Moderately water-cored apples are difficult to sort out visually from sound apples. The inability
to detect such disorders
not only causes a marketing problem butalso prevents investigation of the development of the disorders in intact fruits. Olsen et al. (56) were among the first to
use the differencein optical density at wavelengthsof 760 and8 15 nm to measure
water core concentration in apples. Birth and Olsen (37) made a more definitive
study of this technique to detect water core in Delicious apples. This technique
takes advantageof the physical changesthat occur in apple tissue that affect lightscattering properties. Since the air spaces are eliminated in water-cored tissue. it
scatterslesslightthannormaltissue.Water-coredtissuethustransmitsmuch
more energy. The relative optical density values indicated a water absorption
band at 760 nm. Also, there wererelatively few substances in normal apples that
absorbed energy at about 800 nm, so it was considered the best wavelength rein
gion. An optical density difference of AOD (760-840) gave the best results
identifying water-cored apples.
22
Felsenstein and Manor (57) studied certain surface defects of oranges that
are characterized by color change. They reported that within the vicinity of 667
nm, there was a difference of at least 17% in the intensity
of light reflectance
from the surface of a good orange and one with a blemish. Gaffney (58) investias plug, oleocellosis, rots,
gated fruit grade color and several surface defects such
molds, windscar, scabs, thorn scratches, etc. of oranges and grapefruit. Several
of these defects affect the keeping quality of fruits, whereas others detract from
to be sorted out. The
surface appearance and lower the grade. Such fruits have
Hamlin, Pineapple, and Valencia varieties of oranges and Marsh and Thompson
varieties of grapefruit exhibited definite differences in reflectance characteristics
in the wavelength band of 650-700 nm according to surface color. The wavelength band 550-610 nm was sensitive to changes in surface color or nature due
to various defects.
A color index has been developed to evaluate the quality of orange juice.
The color number (CN) is defined in terms of the tristimulus values X, Y, and
Z as (59):
X
Z
CN = 56.5 - - 18.4 Y
Y
+ 48.2 I
- 8.57
(18)
Applying principal component and discriminant analysis of NIR reflectance

a simple, practical
spectra over a wavelength range of 1100-2498 nm offers
method of detecting 10% pulp wash in orange juice and sugar-acid mixturewith
an accuracy of 90% (60). Visible and UV absorption and fluorescence and emission characteristics of alcoholic solutions of frozen orange concentrates and single strength orange juices can give qualitative detection and quantitative approximation of orange pulp wash in orange juice (61). The absorbance sum at 443,
325, and 280 nm and ratio of absorbance at 443/325nm can provide an estimate
of the percentage total citrus material, orange juice, pulp wash, and dilution
of
the sample. UV-visible absorption and room temperature fluorescence excitation
and emission spectra have been adopted as the official first action to detect adulteration of Florida orange juice with pulp wash (62).
NIR spectroscopy givesa good idea of fruit content, particularly with strawberry jams, for which peaks are obtained
at 770 nm and 1090 nm. FTIR can
distinguish the fruit type in fruit purees (63). It can also detect whether fresh or
freeze-thawed fruit was used to make puree, the level of ripeness in some cases
(e.g., raspberry but not strawberry), fruit variety (e.g., in apples), and any added
sulfur dioxide (64). An FTIR method to determine fruit content of jam has been
reported (65). The FTIR spectra can reliably and reproducibly distinguish between jams of differing fruit content. Furthermore, the spectra are characteristics
of fruit and can act as fingerprints for different fruit types. These methods, therefore, have good potential to verify product authenticity and
to detect adulteration.
Optical Methods
23
The AOD (810-710) was suggested as a nondestructive criterion to detect

in the vicinity
hollow heart disease in potatoes based on the brown substances
of the void (66). This measurement is also capable of indicating other potato
discolorations such as black spots and greening. Porteous et ai. (67) identified
diffuse reflectance at wavelengths between 590 and 890 nm and the bands near
1 100 and 1400 nm as being the most significantin detecting a number of diseases
and defects such as bacterial soft rot, blight, common scab, dry
rot. gangrene,
in
greening, and skin spot. The wavelengths suggested to detect these defects
their order of importance are 650, 710, 1410, 630, 750, and 830nm. In a similar
investigation, Muir et ai. (39) observed that diseased tubers have progressively
reduced diffuse reflectance for several diseases
of potatoes at shorter wavelengths
up to about 800 nm and increased with wavelengths greater than 1100 nm due
to water absorption. The wavelength bands between 590
and 750 nm and the
bands near 950, 1150, 1350, 1470, and 1850 nm were found useful in detecting
various diseases studied.
Mechanical methods for separating potatoes from other materials are
not
very successful because they are similar in their mechanical properties. Differences in reflectance of light and IR radiation by potatoes, stones, and soil clods
offer a possible way of separating them. Palmer (68) reported a red-to-blue ratio,
Rxlxr
,,)-91Xl/R32s
,,, as a very successful criterion in differentiating potatoes from
soil clods. This ratio was found to be unaffected by the size of the object, its
distance from the sensors, and glass. Virtually perfect sorting efficiency was not
unusual. While verifying these results, Story (69) also included the IR radiation
properties to separate potatoes from stones and soil clods. The results indicated
a higher reflectance over the 600that potatoes, like other plant material, showed
1300 nm region and lower reflectance outside this region. Accordingly, the ratio
of the reflectance in the 600- 1300nm region to that in the 1500-2400 nm region
was found to be a more distinctive indicator than the red-to-blue ratio used by
Palmer (68).
2. Food Grains
Separating foreign material such weed
as seeds from thatof other grain cropsis an
important operation in the grain industry. Hawk et al. (70) studied the reflectance
characteristics of 12 grains. Their results indicated that the difference in reflectance between grains in the IR region is small and the greatest differences occur
between 450 and 750 nm. The different reflectance properties were
used in evaluating grain samples for admixture grain, i.e., grains other than the primary one.
1 ) used light reflectance measurementsto detect exterGunasekaran et al. (7
nal cracks in individual kernels of corn. Using a laser light (632.8 nm), they
could detect cracks smaller than 1 mm. The defect detection accuracy was 100%
for broken, chipped, and starch-cracked kernels and 80% for surface-split kernels.
24
Johnson (72) used an absorbance difference measurement (AX(X,-Aq3(,) to determine

heat-damaged,
sprouted,
frosted,
badly
ground-damaged,
and
badly
weather-damaged yellow corn kernels. Though larger differences between damage groups were recorded in the 650-750 nm region than in the 800-1000 nm
region, the latter one was chosen for measurement to minimize effects of natural
color differences in corn kernels.
Birth (73) observed that the slope of the transmittance curves between 750
and 1000 nm was in direct proportion to the amount of smut on wheat samples.
Therefore, excluding the water absorptionband at 970 nm, optical density differences at any two wavelengths in the above range could be indicativeof the total
smut content. The actual smut spore content gave a correlation of 95% with the
measurement of AOD (800-930).
The degree of milling is one of the principal factors in determining the
grade of rice. It is a measure of the extent to which germ and bran layers have
been removed from the endosperm. Extensive milling
is required for complete
removal of the germ and bran layers, which would consequently result
in an
increased percentage of broken kernels. Stermer et al. (74) reported the spectral
transmission of rice with various degrees of milling. Greater changes were observed at approximately 660 and 850 nm. Accordingly, the ratio measurement,
Txso/Tohr,,
was suggested as the criterion. In another method, use was madeof the
fact that protein as well as oil is primarily located in the outer layers of rice to
establish a criterion indicative
of the degreeof milling of rice (75).The oil absorption bands at 928, 1215, and 1725 nm highly correlated with surface and total
lipids. A higher correlation with total lipids ( r = -0.85) than with surface lipids
( r = -0.58) was observed, presumably because1R energy penetrates rice kernels
sufficiently to be absorbed by all the lipids present in the kernel. Beerwinkle and
Stermer (76) utilized this translucence difference between normal and abnornlal
kernels to sort milled rice optically. With this feature included, the efficiency of
a conventional rice sorter was improved
by 50-70%. Stermer (77) developed
objective measurements of the color of milled white rice, which could be used
as an indicator of rice grade, the degree of parboiling, and the extent of starch
gelatinization in parboiled rice.
3. Animal Food Products

Unlike fruits and vegetables where pigments are the dominant factors influencing
their appearance. several factors affect the spectral properties of meat products.
Apart from pigments (predominantly myoglobin), factors
like cellular structure,
surface roughness, and homogeneity can equally affect the appearance
of meat
samples ( 15.78). Forexample, McDougall(78) observedno difference in pigment
or PSE, and normal)
concentration between two qualities (pale-soft-exudative,
of pork muscle which appear different. He also found that the myoglobin and
Optical
25
hemoglobin selectively absorb light, while structural and myofibrillar protein absorb relatively less light but cause more scattering. This suggests that the general
optical quality standards for meat products should take factors other than pigments into consideration.
Davis et al. (79) reported that the interaction between light and muscle
pigments could provide a nondestructive means of evaluating pork muscle quality. They investigated the reflectance spectra of longissimus muscle from pork
loins of the qualities PSE, normal, and dark, firm, and dry (DFD). A pork quality
at 633 and
index (PQI) was suggested based on light reflectance measurement
627 nm as follows:
PQI
- 1.67 - 254log
(RI)
(RI,)
+ 258log
This yielded a correlation of 0.8 with visual rating of quality and 0.86 with the
that the measurements involving
OD of Hart extract. Davis et al. (79) commented
pigments might be affected by the chemical reactions involving porcine myoglobin and by other external factors such as bacterial growth and oxidation. The
light-scattering property of muscle, independent of the above factors, was suget (36) reported
gested as a desirable factor in evaluating muscle quality. Birth al.
a high correlation between the scatter coefficient at 632 nm and the OD of Hart
extract.
The presence of blood and meat spots is one of the most common defects
found in eggs. Its incidence may range from less than 1% to nearly 100% (80).
One of the earliest attempts to develop a spectrophotometric technique to detect
blood in eggs is credited to Dooley(SI), who developed a device to automatically
detect eggs containing blood using radiation
in the region of 1260-1400 nm.
However, Brant et al. (80) identified three absorption bands
for blood at 415,
541, and 575 nm in the visible region. Their method of detecting blood spots i n
white shell eggs was based on the relative transmittance measurement between
555 and 565 nm. Although a success rate of 97.5% was reported, this measurement was specific to the color of the eggshell. Norris and Rowan (82) applied a
similar technique to detect blood spots regardless of shell color. Based on the
relative absorbance measurements at 577 and 600 nm, they could detect 70% of
all eggs having blood spots from 3 to 6 mnl in diameter and 100% having spots
larger than 6 mm in diameter.
C. Composition Analyses
Composition analyses of food materials are very important as quality indices for
a variety of food materials. Such evaluation is normally performed by NIR and
FTIR spectroscopy.
26
7.
MoistureContent
The concept of direct spectrophotometric measurement of moisture content of

food grains was introduced by Norris and Hart (83). In the initial development
diffusetransmittancewasused.However,thediffusereflectancetechniqueis
more popular and is now an accepted technique for rapid analysis of grains and
oilseeds. IR absorption spectroscopy is one of the most versatile methods of determiningmoisture in a variety of substances-gases,liquids,andsolids.By
employing suitable wavelengthsat which maximum absorptionis expected, fairly
reliable, repeatable measurements can be made. The relative advantages and disadvantages of some commonly used moisture determination methods are compared in Table 2. 1R spectroscopy for cereal grains has been investigated in the
700-2400 nm region. Absorption bands of 970, 11 80, 1450, and 1940 nm have
been observed. The moisture content is estimated by comparing the depth of the
band of interest with that for the standard concentration of water (85).
Reflectance and transmittance of grain samples donot change greatly with
moisturecontent(86).Nevertheless,thewaterabsorptionbandsat760,970,
I 180, 1450, and 1940 nm were investigated for spectrophotometric measurement
ofgrainmoisturecontent(87).
The measurementcriterion AOD (970-900)
closely predicted moisture contentof ground wheat samples. In general, it should
be possible to measure the moisture content of a wide range of materials using
the absorption band at 1940 nm on a uniform, thin sample 1-3 mm thick without
any interfering bands. However, for moisture contents greater than 20%, the absorption at 1940 nm is difficult to measure. Hence the 970 nm band should give
greater accuracy. The moisture content of whole, intact peanut cotyledons was
also spectrophotometrically determined (87), and OD (970-900) predicted the
peanut moisture content within ?0.7%.
2. Lipids/Fats
Goulden (88) first used IR radiation to measure fat, protein, and lactose in milk.
Fat measurement was based on absorbance at 1724 cm" (fatA) by ester carbonyl
groups of fat molecules. Protein measurement was based on absorbance at 1538
cm" by peptide bonds of protein molecules, and lactose measurement was based
on absorbance at 1042 cm" by hydroxyl groups of lactose molecules. NIR spectroscopy can be used to analyze moisture, fat, protein, and total solids in cheese
(89,90). Rodriguez-Otero et al. (90)
used NIR reflectance spectroscopyto analyze
fat, protein, and total solids in cheese without any sample treatment.
Norris (91) studied light absorption characteristics of ground beef samples.
Of the observed absorption bands at 540, 575, 640, and 760 nm, he selected the
of ground beef. The
one at 760 nm as a criterion to estimate the fat content
other absorption bands were rejected because they were closely related
to light
absorption by blood. A fat absorption band
at 928 nm was also reportedby Massie
50
Table 2 Advantages and Disadvantages of S o m e Common Moisture Determination Methods

Method
Oven drying
Chemical method. Karl Fisher
Advantages
Standard conventional method
Convenient
Relative speed and precision
Accommodates more samples
Attains desired temperature more rapidly
One of the standard methods
More accurate and precise than other methods
Useful for determining water in oils and fats by preventing samples from oxidizing
Very rapid once apparatus is set up (within minutes)
IR absorption
Can perform multicomponent analysis

Most versatile and selective
Nondestructive
NIR reflectance
Rapid
Precise
Nondestructive
No extraction required
Minimal sample preparation
High sensitivity due to large dielectric constant of water
Convenient to industrial operations with the continuous
measurement system
Universal aplicability
Dielectric capacitance
Source: Ref. 83
Disadvantages
Temperature varies due to particle size, sample moisture.

position in oven, etc.
Difficult to remove bound water
Loss of volatiles
Decomposition of sample (e.g., sugar)
Chemicals of higher purity should be used to prepare reagents
Titration endpoint may be difficult to determine
Reagent is unstable and should be standardized before
use
Titration apparatus should be protected from atmospheric
moisture due to extreme sensitivity of reagent to moisture
Accurate on calibration against reference standard
Dependent on temperature
Dependent on homogenizing efficiency of sample
Absorption band of water is not specific
Reflectance data affected by particle size, shape, packing
density, and homogeneity
Hydroxyl group interferes with amine group
Temperature dependent
Equipment is expensive
Affected by sample texture, packing, mineral content,
temperature, moisture distribution, and acid salts
Calibration difficult far beyond sample pH 2.7-6.7
Difficult to measure bound water at high frequencies
5
0
Q
u)
28
(92).Fromthespectralreflectancedata,
formula:
heproposedthefollowingempirical
where A andBareconstants.Comparisons
of fatcontentestimated bythis
methodwiththevaluesobtained
by chemicalanalysis(Soxhletprocedure)
-+ 1.98% fat.
yielded a correlation of 0.82, and the measurements are within
NIR spectroscopy has been used
to determine the sum of dimer and polymer
triglycerides and acid value to evaluate frying oils (93). This
is a rapid, lowcost technique to assess whether a sample complies with food legislation. FTIR
spectroscopy is another interesting approach to authenticate extra virgin olive oil
(94). The problems associated with identifying adulterated oils are complicated
by the ever-changing nature of adulteration techniques and the numberof procedures that can pass undetected through official quality control.
Some oils can be added to other oils without being detected
by routine
physical and chemical characteristics due to their fatty acid composition. In such
cases, gas-liquid chromatography (GLC) analysis of fatty acids proves useful. IR
spectra between 300 and 357 cm-l and 770to 1175 cm" can distinguish between
oils of peanut, sesame, sunflower, etc. Peanutoil has a characteristic band at 9 13
cm",sunfloweroilat847,andsesameoilat812and
913 cm". IR spectra
oil and its
between 4000 and 850 c1n-l have shown differences between olive
adulterant, rapeseed oil. Differences have been noted at 3100 and 1750 cm" and
of the differential spectrum
from 1400 to 1300 cm", the most striking feature
being the negative peaks at 1130 and 1080 cm" with a characteristic contour
from 900 to 1200 cm". Thesecharacteristicspersist in a mixture containing
in
as little as 10% rapeseed oil. These differences are attributed to differences
unsaturated fatty acids, particularly oleic and linoleic acids (95).
In contrast to NIR, FTIR has much to offer the analyst because specific
bands may be assigned to specific chemical entities. Statistical correlation methods are not always necessary, but they are not excluded and may be required
in
very complicated mixtures (63). This techniques has been widely used to deter(96), meat (97), sweetened condensed
mine fat, moisture, and protein in butter
milk (98), and other high-fat products (99). It has also been used to monitor the
oxidation of edible oils (100) and to determine the level of tram-unsaturation in
fat (101).
By combining attenuated total reflectance and mid-IR spectroscopy with
statistical multidimensional techniques, Safer et al. (102) obtained relevant information from mid-IR spectraof lipid-rich food products. Wavelength assignments
for typical functional groups in fatty acids were made for standard fatty acids.
Optical Methods
29
Absorption bands around 1745 cm" due to carbonyl group, 2853 and 2954 cm"
due to C-H stretch, 3005 and 960 cm" due to
C = C bonds, 1160 cm" due
to C - 0 bonds, and 3450 and 1640 cm" due to 0 - H bonds were observed.
Water strongly absorbs in the region of 3600-3000 cm-l and at 1650 cm" in
butter and margarine, allowing one to rapidly differentiate them as a function of
their water content. Principal component analysis was used
to emphasize the
difference between spectra and to rapidly classify
27 commercial samples of oils,
butter, and margarine.
Belton et al. (103) studied the components of fat, protein, and sugar in
confectionery products usingFTIR spectroscopy coupled with photoacoustic and
attenuated totalreflectancedetectionmethods.Theyconcluded
that peaks at
1744, 1477- 1400, 1240, and 1 195- 1 129 cm" could be from an ester carbonyl
group, C-H bond, and C - 0 stretching of fat, respectively. Peaks at 1650 and
1540 cm" are from protein,
and those at 1128 to 952 cm" are from sugars.
3000 cm" and at 1650 cm" in
Water is strongly absorbed between 3600 and
fat-rich foods (96,103). Principal component analysiswas used to emphasize the
differences between spectra and to rapidly classify each sample (96).
Usually wavelength assignments for typical functional groupsin fatty acids
A) forestercarbonylgroups
areabsorptionbandsaround1745cm"(fat
in methylene
(R(CO)OR/OH), 2930 and 2853 cm"' (fat B) for C-H stretch
groups, and 1 160 cm" for C - 0 bonds of lipid (104).
IR spectroscopy hasalso been used to detect adulterationof fat with lowerquality/cost oil. Attempts have been made to detect concentrations of less than
10% of vegetable or animal fat
in butter fat by GLC in conjunction with IR
spectroscopy (percentage transmission at 967 and 948 cm", denoted as T967
and T948 and assigned to isolated trans and conjugated cis-trans isomers) can
reliably distinguish butter from its adulterant substitute fats (95).
3. ProteinContent
Proteins have three characteristic absorbance bands
in the mid-IR spectrum (104).
Two of these, amide I (about 1600- 1700 cm") and amide 111 (about 1200- 1400
cm"), are sensitive to polypeptide backbone conformation and might be able to
I band is moreintense, but it
distinguishbetweenproteins(105).Theamide
overlaps with an intense water deformation band at 1645 cm". The amide 111
band, although less intense, is not overlapped by water absorptions. This band
has been used as a tool to detect adulterations of NDM with SPC (106).
Wheat protein content and grain hardness can be rapidly determined
by
IR and NIR spectroscopy (107). The NIR and Brabender hardness tester results
correlate significantly with percentage of dark hard and vitreous grains as shown
by commercial red winter wheats which have similar protein contents. NIR spec-
30
troscopy has also been used to differentiate between hard red winter and hard
red spring wheat. Examination of the principal component factors has indicated
that hardness, protein level, and the interaction of water with protein and other
constituents are responsible for correct classification based on NIR (108).
IR absorption can be used to determine the protein content in whole milk
at 6460 nm, which is the absorption maximum for the peptide bond. Other components of milk such as lactose and fat can be simultaneously measured at their
respective absorption bands at 5730 and 9597 nm. Water absorbs significantly
at 1000-5000 cm. Therefore, it interferes excessively with protein absorption
bands in the IR spectrum. For protein determination in milk, this couldbe alleviated by using a double-beam spectrophotometer with water in the reference cell
and milk in the sample cell (109).
NIR spectroscopy is a popular method for determining protein
in cereal
products primarily due to its speed, simplicity of operation, safety, and low operating cost. To avoid excessive interference by starch, fat, and water, a wavelength of 4590 cm corresponding to a combined vibration of amide groups is
chosen to quantify protein components ( 1 IO).
VI.
SUMMARYAND FUTURE TRENDS
Optical propertiesof food and biological materials vary widely and are dependent
upon many factors. Quality evaluation based
on these properties requires accurate
optical property values. Undera given setof conditions, precise values can probably be obtained for any particular substance by careful measurements. However,
good estimates can be made for many materials on the basis of the investigations
already reported. Such optical property estimates will permit a quick evaluation
of various properties that have servedas quality indices and help in the selection
of those most suitable for any particular application.
Generally, nondestructive quality evaluationof food and biological materials focuses on three major aspects: maturity and/or ripeness evaluation, internal
For maturity evaluation,
and external defect detection, and composition analysis.
the interaction of various pigments and the changes associated with them during
maturation are taken as the prime indicator. The importance of chlorophyll in
this regard has been adequately established. Both external and internal defects
have been found to affect the normal propertiesof interaction of light with products in consistent ways. Hence, defect detection is accomplished by comparing
optical propertiesof a normal product with thoseof the defective ones. Moisture,
protein, fat or oil content, and other compositions have been analyzed based on
certain absorption bands in the electromagnetic spectrum.
The wide variety of sizes, shapes, and textures of food products makes
most commercial instruments difficult to use for measuring optical characteristics
Optical
31
since special geometric designs are needed. With technological advancements

such as fiber optics, laser, etc., this problem has been partially overcome. Fiber
optic technology has the advantage of detecting light of extremely low intensity
by providing a good light seal and by eliminating the effects of variations such
as fruit size and distance from the light source. This has also been found to be
very useful in high-speed operations. Low-power lasers can be of great help because the laser beam is highly directional, coherent,very bright, and has a welldefined beam diameter that can be focused down
to small sizes to detect very
local defects.
The spectrophotometric techniques so far investigated rely heavily on ema better engineering approach, interaction
pirical data and statistical analyses. For
of lightwithagriculturalproductsshouldbeanalyzedanalytically.Chenand
to
Nattovetty (1 11) indicated that if a mathematical model can be developed
represent the distribution of diffused light in fruit, the effectsof various parameters could be studied more thoroughly. Many food products may be considered
as turbid or translucent, i.e., the incident light energy does not traverse the object
in a rectilinear manner but is scattered from its original direction of travel. Thus,
not only the absorption
of energy within the sample, but also the scatter and
changes in the direction of travelof energy within the sample must be considered.
Mathematical description of scatter, diffuse reflectance, and diffuse transmittance
are extremely valuable in providing a means of obtaining insight into the interaction between light and biological materials where multiple scattering predominates.
The NIR and FTIR techniques are finding significantly increased usein the
food industry for quality evaluation and control applications. The ability of such
a variety of foods have made
techniques to provide compositional data rapidly for
possible some on-line quality evaluationsthat were traditionally done off-line in
a laboratorysetting.Determination of adulteratedentities in manyliquidand
in computer and
powdered foods is a very good example. With new advances
to findevenwider
opticaltechnology,theNIRandFTIRmethodsarelikely
applications. The optical methods will also become more commonplace in rapid
microbial testing for improved food safety.
REFERENCES
A Kramer, BA Twigg. Quality Control
for the Food Industry. Vol.1. Fundamentals.
Westport, CT: AVI Publishing Co., 1970.
2. FJ Francis, FM Clydesdale. Food Colorimetry: Theory and Applications. Westport,
CT: AVI Publishing Co., 1975.
3 . JB Hutchings. Food Color and Appearance. 2nded. Gaithersburg, MD: Aspen PublishersInc..1999.
1.
32
4.
5.
6.
7.
X.
9.
10.
I 1.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
RC Gonzalez, RE Woods. Digital Image Processing. Reading, MA: Addison-Wesley Publishing Company, 1992.
R Chizzolini, E Novelli, A Badiani, P Rosa, G Delbono. Objective measurements
34(1):49-77, 1993.
of pork quality: evaluation of various techniques. Meat Sci
R Chizzolini, E Novelli, A Badiani,
G Delbono, P Rosa. Objective evaluation of
pork quality: results of on-line measurements. Meat Sci
34(1):79-93, 1993.
GS Birth,GLZachariah.Spectrophotometry
of agriculturalproducts.Trans Am
Soc Agric Eng 16548, 1973.
APolesello,RGiangiacomo.Application
of nearinfraredspectrophotometryto
the nondestructive analysis of foods: a review of experimental results. CRC Crit
RevFoodSciNutr 18:203,1983.
GS Birth.Electromagneticradiation:optical.In:
ZA Henry,ed.Instrumentation
and Measurement for Environmental Sciences. St. Joseph, MI: American Society
of Agricultural Engineering Special Publication No.
SP-0375, 1975.
GS Birth. The light scattering properties of foods. J Food Sci 43:915,1978.
FM Clydesdale. Calorimetry-methodology and applications. CRC Crit Rev Food
SciNutr 10:243,1978.
BG Osborne,TFearn.NearInfraredSpectroscopy
in FoodAnalysis.Avon,England: Longman Scientific and Technical, 1986.
KH Norris. Multivariate analysis ofnew materials. In: LW Shemit, ed. Chemistry
and World Food Supplies: The New Frontiers. CHEMRAWN 11. Oxford, England:
Pergamon Press, 1982, p 527.
G Kortum.ReflectanceSpectroscopy:Principles,Methods,Applications.New
York:Springer-Verlag, 1969.
GS Birth. How light interacts with foods. In: JJ Gaffney, ed. Quality Detection
in
Foods. St. Joseph, MI: American Society of Agricultural Engineering; 1976, p6.
P Chen. Use of optical properties of food materials in quality evaluation and materials sorting. J Food Proc Eng 2:307, 1978.
PA Suci, JD Vrany, MW Mittelman. Investigation of interactions between antimicrobial agents and bacterial biofilms using attenuated total refection Fourier transform infrared spectroscopy. Biomaterials 19:327-339,1998.
DA Skoog, FJ Holler, TA Nieman. In: Principles of Instrumental Analysis. 5th ed.
Philadelphia: Harcourt Brace College Publishers, 1998, pp 143-191.
RH Wilson, EK Kemsley. Infrared spectroscopic methods. In: AC Pinder,
G Godfrey, eds. Food Process Monitoring Systems. London: Blackie Academic and Professional, 1993, p 40.
PR Griffiths,JAdeHaseth.In:FourierTransformInfraredSpectroscopy.New
York: Wiley-Interscience Publications, 1986.
J Larrabee, S Choi.Fouriertransforminfraredspectroscopy.MethodsEnzymol
226:289-305,1993.
FRvan de Voort, AA Ismail. Proximate analysis of foods by mid-FTIR spectroscopy. Trends Food Sci Techno1 2: 13- 17, 199 1.
23. FR van de Voort. Fourier transform infrared spectroscopy applied to food analysis.
Food Res Int 25:397-403, 1992.
24. NS Letzelter,RHWilscon,ADJones,
G Sinnaeve. Quantitative determination of
22.
Optical Methods
25.
26.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
33
the composition of individual pea seeds

by Fourier transform infrared photoacoustic
spectroscopy. J Sci Food Agric 67:239-245, 1995.
G Downey. Food and food ingredient authentication of mid-infrared spectroscopy
and chemometrics. Trends Anal Chem 17:418-424, 1998.
SV Compton, DAC Compton. Optimization of data record by internal reflectance
spectroscopy.In: PBColeman, ed.PracticalSamplingTechniquesforInfrared
Analysis. Boca Raton, FL: CRC Press, Inc., 1993, pp 55-92.
L Rintoul, H Panayiotou,S Kokot, G George, G Cash, R Frost, T Bui, P Fredericks.
Fourier transform infrared spectrometry: a versatile technique for real world samples.Analyst123:571-577,1998.
MW Urban. Surfactantsin latex films; interactions
with latex components and quantitative analysis using ATR and step-scan PAS FT-IR spectroscopy. Progress Organic Coatings 32:215-229, 1997.
BC Smith. Fundamentalsof Fourier Transform Infrared Spectroscopy. Boca Raton,
FL: CRC Press, 1996.
The Complete Guide to IT-IR. Shelton, CT: Spectra-Tech, Inc., 1998.
M Blanco,J Coelle, A Eustaquio,H Iturriaga, S Maspoch. Development and validation of a method for the analysis of a pharmaceutical preparation by near-infrared
diffuse reflectance spectroscopy. J Pharm Sci 88:551-556, 1999.
JB Reeves,111, CM Zapf. Mid-infrared diffuse reflectance spectroscopy for discriminant analysis of food ingredients. J Agric Food Chem 463614, 1998.
JB Reeves, 111. Near-versus mid-infrared diffuse reflectance spectroscopy for the
quantitative determination of the composition of forages and byproducts. J Near
Infrared Spectrosc 2:49-57, 1994.
R Braindet, EK Kemsley, RH Wilson. Discrimination of Arabica and Robusta in
instantcoffee by Fouriertransforminfraredspectroscopyandchemometrics.
J
Agric Food Chem 44: 170- 174. 1996.
G Downey, P Robert, D Bertrand, RH Wilson, EK Kemsley. Near- and mid-infrared
spectroscopics in food authentication: coffee varietal identification. J Agric Food
Chem 45:4357-4361, 1997.
DR Bittner. KH Norris. Optical properties of selected fruits vs. maturity. Trans
Am
Soc Agric Eng 1 1:534, 1968.
GS Birth, CE Davis, WE Townsend. The scatter coefficient as a measure of pork
quality. J Animal Sci 46:639, 1978.
GS Birth, KL Olsen. Nondestructive detection of water core in Delicious apples.
Proc Am Soc Hort Sci 85:74, 1964.
GS Birth, KH Norris, JN Yeatman. Nondestructive measurement of internal color
of tomatoes by spectral transmission. Food Technol 11:552, 1957.
AV Muir, RL Porteous, RL Wastie. Experiments in the detection of incipient diseases in potato tubers by optical methods. J Agric Eng Res 27: 131, 1982.
J N Yeatman, KH Norris. Evaluating internal quality of apples with new automatic
fruit sorter. Food Technol 19:123, 1965.
AP Sidwell, GS Birth, JV Ernest, C Golumbic. The use of light transmittance techniquestoestimatethechlorophyllcontentandstageofmaturationofElberta
peaches.FoodTechnol19:123,1961.
34
42. S Gunasekaran, MR Paulsen, GC Shove. Optical methods for nondestructive quality evaluation of agricultural and biological materials. J Agric Eng Res 32(3):209,
1985.
43. B Zion, P Chen, MJ McCartney. J Sci Food Agric 67:423-429, 1995.
44. V Bellon, SI Cho, GW Krutz, A Davenel. Food Control 3(1):45-48, 1992.
45. V Bellon, G Rabatel, C Guizard. Food Control 3( l):49-54, 1992.
46. AH Kramer. JE Gates, KD Demoree, AP Sidwell. Spectrophotometric investigations on peanuts with particular reference to estimation of maturity. Food Technol
17(8):90,1963.
47. EO Beasley,JW Dickens. Light transmissionof peanut oilas an objective measurement related to quality of raw peanuts. ASAE Paper No. 67-809, American Society
of Agricultural Engineers, St. Joseph, MI, 1967.
48. A Nussinovitch, G Ward,E Mey-Tal. Gloss of fruit and vegetables. Lebensm Wiss
Technol29(1):184-186,1996.
49. G Ward, A Nussinovitch. Gloss properties and surface morphology relationships
of fruits. J Food Sci 61(5):973-977, 1996.
SO. G Ward, A. Nussinovitch. Peel gloss as a potential indicator of banana ripeness.
Lebensm Wiss Technol 29:289-294, 1996.
51. M Ingle, JF Hyde. The effect of bruising on discoloration and concentration of
phenoleic compounds in apple tissue. Proc Am SOC Hort Sci 93:738, 1968.
by leaves. Plant Physiol 47:
52. JT Woolley. Reflectance and transmittance of light
656,1971.
53. GK Brown, LJ Segerlind, R Summitt. Near-infrared reflectance of bruised apples.
Trans Am SOC Agric Eng 17: 17, 1974.
54. WK Bilanski, CL Pen,DR Fuzzen. Apple bruise detection using optical reflectance
parameters. Can Agric Eng 26(2): 1 1 1-1 14, 1984.
55. WS Reid. Optical detection of apple skin, bruise, flesh, stem, and calyx. J Agric
Eng Res 21:291, 1976.
56. KL Olsen, HA Schomer,GS Birth. Detection and evaluationof water core in apples
by light transmittance. Washington State Hort Assoc Proc 58:195, 1962.
57. G Felesenstein, G Manor. Feasibility study into the development of an improved
photoelectric device for sorting citrus fruits for surface defects. Trans
An1 SOC Agric
Eng 16: 1006, 1973.
58. EE Gaffney. Reflectance properties of citrus fruits. Trans Am SOC Agric Eng 16:
310. 1973.
of Food Analysis. Vol.
59. C Calvo. Optical properties. In: LML Nollet, ed. Handbook
I . New York: Marcel Dekker, 1996.
60. M Twomey,G Downey, PB McNulty.The potential of NIR spectroscopy for detection of the adulteration of orange juice. J Sci Food Agric 67(1):77-84, 1995.
61. DR Petrus, NA Dunham. Methods for detection of adulteration in processed citrus
products. In: S Nagy, JA Attaway, eds. Citrus Nutrition and Quality. ACS Symposium Series 143, 1980, pp 395-421.
62. DR Petrus, JA Attaway. J Assn Official Anal Chem 68(6):1202-1206. 1985.
63. PS Belton, AM Saffa, RH Wilson. Use of Fourier transform infrared spectroscopy
for quantitative analysis:a comparative study for different detection methods. Analyst 112:1117-1120,1987.
Optical Methods
35
64. M Deferenz.EK Kemsley, RH Wilson. Use of infrared spectroscopy and chemometrics for the authentication of fruit purees.
J Agric Food Chem 43( I):l09-113,
1995.
65. RH Wilson. Fourier transform mid-infrared spectroscopy for food analysis. Trends
Anal Chem 9(4): 127- 13 I , 1990.
66. GS Birth. A nondestructive technique for detecting internal discoloration in potatoes. Am Potato J 3753, 1960.
The identification of diseases and defects in
67. RL Porteous, AV Muir, RL Wastie.
potato tubers from measurements of optical spectral reflectance. J Agric Eng Res.
26:151,1981.
68. J Palmer. Electronic sortingof potatoes, and clods by their reflectance.J Agric Eng
Res6:104,1961.
69. AG Story. Spectral reflectance of light and infrared radiation by potatoes, stones,
andsoilclods.In:JJGaffney,
ed. QualityDetection in Foods. St. Joseph,MI:
American Society of Agricultural Engineers, 1976, p
83.
70. ALHawk,HHKaufmann,CAWatson.Reflectancecharacteristicsofvarious
grains. ASAE Paper No. 69-357, American Society of Agricultural Engineers, St.
Joseph, MI, 1969.
71. S Gunasekaran, MR Paulsen, GC Shove. A laser optical system for detecting corn
kernel defects. Trans Am Soc Agric Eng 29( 1):294-298, 304, 1986.
72. RM Johnson. Determining damagein yellow corn. Cereal Sci Today 7(
I ) : 14, 1962.
13. GS Birth. Measuring smut content of wheat. Trans Am Soc Agric Eng Res 3:19,
1960.
74. RA Stermer, HW Schroeder, AW Hartstack, CH Kingsolver. A rice photometer
for measuring the degree of milling of rice. Rice J 67(5):24, 1962.
75. RA Stermer, CA Watson, E Dikeman. Infrared spectra of milled rice. ASAE Paper
No.76-3030,AmericanSocietyofAgriculturalEngineers,
St. Joseph,MI,
1976.
76. RBeenvinkle, RA Stermer.Adevicetofacilitateopticalsortingofmilledrice
based on translucence difference. ASAE Paper No. 71-373, American Society of
Agricultural Engineers, St. Joseph, MI. 197 I .
77. RA Stermer. An instrument for objective measurement of degree of milling and
color of milled rice. Cereal Chem 45:358. 1968.
78. DB McDougall. Characteristics of appearance of meat.1. The luminous absorption,
scatter, and internal transmittance of the lean of bacon manufactured from normal
and pale pork. J Sci Food Agric 21568, 1970.
79. CE Davis, GS Birth, WE Townsend. Analysis of spectral reflectance for measuring
pork quality. J Animal Sci 46:634, 1978.
80. AW Brant, KH Norris, G Chin. A spectrophotometric method for detecting blood
in white-shell eggs. Poultry Sci 32:357, 1953.
81, WD Dooley. Method and apparatus for detecting the presence of blood in an egg.
U.S. Patent 2,321,899 (1943).
82. KH Norris, JD Rowan. Automatic detection of blood in eggs. Agric Eng 43(3):
154.1962.
83 KH Norris, JR Hart. Direct spectrophotometric determination of moisture content
of grain and seeds. In: A Wexler, ed. Humidity and Moisture: Measurement and
36
ControlinScienceandIndustry.NewYork:ReinholdPublishingCorporation,
1965.
84. YW Park. Determination of moisture and ash contents
of food.In:LMLNollet,
ed. Handbook of Food Analysis. Vol.
1. New York: Marcel Dekker, 1996.
85. EKarmas.FoodTechno134(1):52,1980.
86. DR Massie, KH Norris. Spectral reflectance and transmittance properties of grain
in the visible and near-infrared. Trans Am SOC Agric Eng
8598, 1965.
87. KH Norris, JR Hart. Direct spectrophotometric determination of moisture content
of grains and seeds. Proceedings of 1963 International Symposium on Humidity
and Moisture, Principles and Methods
of Measuring Moisture in Liquid and Solids.
Vol. 4. New York: Reinhold Publ., 1963, p 19.
88. JDS Goulden. Quantitative analysis of milk and other emulsions
by infrared absorption.Nature191:905-912,1961.
89. MM Pierce, RL Wehling. Comparison of sample handling and data treatment methods for determining moisture and fat in Cheddar cheeseby near-infrared spectroscopy. J Agric Food Chem 42:2831-2835, 1994.
90. JL Rodriguez-Otero, M Hermida, A Cepeda. Determination of fat, protein and total
solids in cheese by near-infraredreflectancespectroscopy. J AssnOfficialAnal
Chem Int 8:802-806, 1995.
91. KH Norris.Measuringlighttransmittancepropertiesofagriculturalcommodities.
Agric Eng 39:640, 1958.
92. DR Massie. Fat measurementof ground beef with a gallium arsenide infrared emitter.ASAEPaperNo.73-6503,AmericanSocietyofAgriculturalEngineers,
St.
Joseph, MI, 1973.
93. AJ Boot, AJ Speck. J Assn Official Anal Chem Int 77(5): 1 184-1 189, 1994-95.
94. YW Lai, EK Kemsley, RH Wilson. Quantitative analysis of potential adulterants of
extra virgin olive oil using infrared spectroscopy. Food Chem 53(1):95-98, 1995.
95.RSSinghal,
PR Kulkarni, DV Rege.HandbookofIndicesofFoodQualityand
Authenticity. Cambridge, England: Woodheard Publishing Ltd., 1997.
96. FR van de Voort, J Sedman, G Emo. A rapid FTIR quality control method for fat
and moisture determination in butter. Food Res Int 25:193-198, 1992.
97. B Dion, M Ruzbie, FR van de Voort, AA Ismail, .IS Blais. Determination of protein
and fat in meat by transmission Fourier transform infrared spectrometry. Analyst
119:1765-1771,1994.
98. N Nathier-Dufour, J Sedman, FR van de Voort. A rapid ATR/FTIR quality control
method for the determinationof fat and solidsin sweetened condensed milk. Milchwissenschaft 50:462-466, 1995.
99. FRvan de Voort, J Sedman, AA Ismail. A rapid FTIR quality-control method for
determining fat and moisture in high-fat products. Food Chem 48:213-221.
100. FR van de Voort, AA Ismail, J Sedman, G Emo. Monitoring the oxidation of edible
oils by Fourier transform infrared spectroscopy. J Am Oil Chem SOC 71:243-253,
1994.
101. F Ulberth, HJ Haider. Determination of low level trans-unsaturation infats by Fourier transform infrared spectroscopy. J Food Sci 57:1444-1447, 1992.
102. M Safer, D Bertrand, P Robert, MF Devaux, C Genot. Characterization of edible
Optical Methods
103.
104.
105.
106.
107.
108.
109.
1 10.
111.
37
oils, butters and margarinesby Fourier transform infrared spectroscopy with attenuated total reflectance. J Am Oil Chem Soc 71:371-377, 1994.
PS Belton,AM Saffa, RH Wilson. The potential of Fourier transform infrared spectroscopy for the analysis of confectionery products. Food Chem 28:53-61, 1988.
RM Silverstein, FX Webster.In:SpectrometricIdentification
of OrganicCompounds. 6th ed. New York: John Wiley & Sons, Inc., 1998, pp 71-1 1 1 .
MR Nyden, GP Forney, K Chittur. Spectroscopic qualitative analysis of strongly
interactingsystems:humanplasmaproteinmixtures.ApplSpectroscopy42(4):
588-594,1988.
IV Mendenhall, RJ Brown. Fourier transform infrared determination of whey powder in nonfat dry milk. J Dairy Sci 74(9):2896-2900, 1991.
SR Delwiche,G Weaver. Bread quality of wheat flour by near-infrared spectrophotometry: feasibility of modeling. J Food Sci 59(2):410-415, 1994.
SR Delwiche, KH Norris. Classification of hard red winter wheatby near-infrared
diffuse reflectance spectroscopy. Cereal Chem 70( 1):29-35, 1993.
H Guillou,JP Pelissier, R. Grappin. Methods for quantitative determinationof milk
proteins. Le Lait 66(2):143-175, 1986.
JV Camp, A Huyghebaert. Protein. In: LML Nollet, ed. Handbook of Food Analysis. Vol. 1. New York: Marcel Dekker, 1996, p 277.
P Chen. VR Nattovetty.Lighttransmittancethrough
a region of an intact fruit.
ASAE Paper No. 77-3506, American Society of Agricultural Engineers,
St. Joseph,
MI,1977.
This Page Intentionally Left Blank
Computer Vision
Suranjan Panigrahi
1.
INTRODUCTION
Assessment or evaluation of food products is essential for ensuring their quality

and safety. Rising consumer awareness and expectation for high quality along
with strict legislation for food safetyhas necessitated quality evaluationof every
food product being manufactured or processed. Consequently, the development
and/or identification of advanced, reliable, fast, and cost-effective formsof nondestructive sensors and/or sensing techniques arein high demand. Computer vision is a powerful technique to extract and quantify features for food quality
assessment and control. It offers the advantagesof accurate quantification of images and rapid data handling. Computer vision has been a proven technology for
a varietyof nondestructive methods to evaluate food product characteristics ranging from dimensional measurements (length, width, shape, other geometrical attributes) to texture, color, defects and diseases, to three-dimensional analysis of
food quality. With rapid advances in electronic hardware and other associated
computer imaging technologies, the cost-effectiveness of computer vision systems has greatly improved.It is estimated thatby the year 2000, sales of machine
vision systems in the North American food sector will be approximately $148
million (1 ).
39
40 Gunasekaran
11.
and
Panigrahi
COMPUTERVISION SYSTEMS
Computer vision is the science that develops the theoretical and algorithmic basis
by which useful information about an object or scene can be automatically extracted and analyzed from
an observed image, image set,
or image sequence.
Computer vision is also known as machine vision or computer imaging. It is a
branch of artificial intelligence technique and deals with simulating human vision.
1):
The essential components of a typical computer vision system are (Fig.
Computer (analogous to the human brain)
Sensor or camera (analogous to the human eyes)
Illumination system (light source and illumination chamber, etc.) to facilitate image capture
Frame grabber/digitizer to digitize the image information from the camera,
monitor(s)
In todays modern computer imaging system, the camera and frame grabber
are joined together to form the digital camera. The use of digital cameras thus
eliminates theneed to use a framegrabber in the computer. The imageis digitized
in the camera and sent to the computer for further processing. More modern
computer imaging systems do not need a separate monitor either. Images can be
displayed directly on a high-resolution computer monitor. For special applications, many users still use a high-resolution display monitorto separately display
images.
Recognizing and extracting useful object features from image data are complex tasks involving a series of steps that can be grouped into three major parts
Frame Grabber/
Computer
Sensor/
Camera
Light source
Illurnination
Chamber
Fig. 1 Schematic of a typicalcomputervisionsystem.
Display
Monitor
Computer Vision
41
Fig. 2 Basic steps in digital image processing.
(Fig. 2): image acquisition, image processing, and image understanding. Image
acquisition deals with such issueskomponents as illumination, camera, digitizer,
etc. The image processing step encompasses preprocessing, segmentation, and
feature extraction. The image understanding part consists of image recognition
so as to make
and interpretation. Eachof these steps must be carefully performed
each subsequent step progressively easier and result inan improved end result.
For example, a poorly formed or acquired image cannot provide good results
with any amount of further processing.
All the steps are closely linked
with
knowledge base available about the system studied and the featuresof interest.
111.
IMAGE ACQUISITION
A.
Digital Images
A digital image can be defined as a spatial representationof an object or scene.

A digital monochrome image is a two-dimensional (2-D) light-intensity function,
denoted by I(x,y), where the value or amplitude of intensity I at spatial coordinates (x,y) is typically proportional to the radiant energy received in the electroor detector (the camera) is sensitive in a small
magnetic band to which the sensor
area around the point (x,y). As far as the computer is concerned, the image is a
matrix (x,y) of numerical values, each representing a quantized image intensity
value. Each matrix entry is known as a pixel (short for picture element). The
total number of pixels in an image is determined by the size of the 2-D array
used in the camera. Most commonly used cameras have a spatial resolution of
42
Panigrahi and Gunasekaran
5 12 X 480 or 640 X 480. For best results it is important to match the spatial
resolution of the camera to that of the frame grabber.
The intensity of the monochrome image is known as the gray level. The
limit on gray level is that it is positive and finite. The gray level interval (from
low to high) is called a gray scale. A common practice is
to shift this interval
numerically to the interval (0,L) where the lowest value 0 represents pure black
All intermediate values are
and the maximum value L represents pure white.
shades of gray varying continuously from black to white. For example, when an
8-bit integer is used to store each pixel value, gray levels range from
0 to 255
(i.e., 2" - I to 2' - I).
Inferring an object's size, shape, position, orientation, and other attributes
from the spatial distribution of gray levels requires the capability to infer which
pixels belong to the object and which do not. Then, from the pixels that belong
to the object, it requires the capability to identify the object features of interest.
Algorithms have been developed to translate the gray levels of a pixel in a way
that accentuates the desired information.
In the case of color images, the image intensity
is represented by three
components representing red, green, and blue (RGB system) or hue, saturation,
and intensity(HSI system). Further detailsof the color digital image are presented
in a following section.
B. illumination
The prerequisite for any vision application is that the features to be examined
can be seen in the image. Therefore, despite all the progress in image analysis/
processing algorithms, the performance of the camera and illumination subsystem
can greatly affect the reliabilityof a machine vision application. A well-designed
lighting and illumination system can assist in the accuracy and success of image
analysis by enhancing image contrast. Good lighting will improve feature discrimination,reduceprocessingtime,andreduceprocessinghardwarerequirements. Thus, it is almost always cheaper to improve lighting than image processing (2). Food materials are nonhomogeneous and randomly oriented; the raw
materials may be dirty. Singulation of objects for examination is often difficult,
so we have to cope with objects that touch and/or overlap, which
may cause
shading during image acquisition. Therefore, vision applications in the food industry present unusual challenges when designing proper illumination systems.
Illumination consists of selecting appropriate light sources and identifying
suitable configurations for the light sources so as to obtain the highest quality
images. The geometryof the imaging system should be well known. This requirement is especially important for dimension measurements. When the viewing
geometry is more complicated, either becauseof the nonplanar image surface or
nonperpendicular imaging angle, measurements are more
difficult and require
Computer Vision
43
determining the geometryof the imaging system. Elaborate discussions of different types of illumination techniques for general machine vision applications have
been presented by other authors (3,4). Most lighting arrangements can be grouped
as either front-lighting or back-lighting. The front-lighting option is best suited
for obtaining surface characteristics of an object, while back-lighting is best for
(5)
subsurfacefeatures. For example, using back-lighting, Gunasekaran et al.
examined internal stress cracks
in corn kernels and Upchurch and Throop
(6)
detected watercore in apples. The appropriatenessof a well-designed illumination
system can be evaluated by the suitability of acquired images for successful further processing. The most commonly used illuminations system configurations
are summarizedin Fig. 3. Associated advantages and disadvantages of these techniques are compared in Table 1.
A wide variety of light sources and lighting arrangements are available
(2).
Most general computer vision applications are implemented using either incandescent or florescent lighting. However, use of polarizers and polarized light can
improve the light intensity contrast, eliminate unwanted glare, and minimize diffuse reflectance (7). This is especially suitable for transparent and translucent
objects. Since an objects color dependson illumination, color measurements are
easily affected by changesin the color temperatureof an incandescent bulb. Thus,
or colorvalues,requiresa
measuringbrightnessinformation,suchasdensity
very stable illumination source and sensor. Bright specular reflections may cause
saturation, blooming, or shifts in image magnification. Sometimes the color of
two objects will appear similar under one light source but much different under
another. So, a number of light sources of different spectral responses must sometimes be tried when attempting to maximize image contrastfor the best possible
results. For multiple or brightly colored fruits and vegetables, a multiple spectral
lighting system is needed to assure accuracy over a large spectral range. Spectral
reflectance properties of products should be considered when developing an appropriateilluminationsystem(lightingandviewinggeometries,lightsources,
(8). The specand sensor components) to obtain maximum discrimination power
tral output of different light sources can be obtained from respective manufacturers.
For on-line evaluations where speed of operation becomes an important
look the same
criterion, global uniformity (i.e., the same type of feature should
wherever it appears in the image) is essential. This means that brightness and
color values are the same and
thus it requires uniform, consistent image illumination (9). Furthermore, the optomechanical constructionof a camera and illuminator should withstand environmental conditions suchas mechanical vibrations and
dust common in industrial applications. Strobe lightingis useful for on-line applications to virtually arrest themotion to aid in acquiring images without worrying
about image blur due to image motion.
The strobe repetition rate should be
selected to match the speed of object motion. A strobe unit designed for machine
44
Light
Sources
Light
Sources
Diffuser
Diffuser
kpiece
Fig. 3 Different configurations of illumination system for computer vision. (A) Diffuse
front illumination, used for general top lighting. (B) Directional front illumination creates
shadows and will not reflect into the camera if surface
is highly reflective. (C) Light tent
(cloudy day) is nondirectional, totally diffuse top lighting that produces illumination like
that on a cloudy day (good for metal parts and electronic components).
(D) Back-lighting
through a collimating lens
so that the light rays
are pseudo parallel. (E) Dark field illumination in which incident light reflects away from the camera and illumination
is created from
specular reflections. (F) Diffuse back-lighting, in which light is on the opposite side of
the part as the camera and goes through a diffusing material suchas lexan or opal glass.
(G) Low-angle illumination, in which incident lighting is almost parallel to the surface
of the part. (H) Polarized front illumination, involving front-lighting with a polarizer on
the light and a cross-polarizer on the lens.
(I) Polarized back-lighting, in which a polarizer
and a cross-polarizer are on opposite sides of the part over some form of back-lighting.
(J) Stmbed illumination, in which microsecond duration lighting
is used to freeze the
motion of moving parts. (K) Structured light, in which a plane of light generated via
structured white light with focusing optics,
or laser line converter,is used to show contour/
3-D information of the part. (L) Coaxial lighting, in which the illumination is along the
same path as the cameras viewing path. (Courtesy of Machine Vision Association, Society
of Manufacturing Engineers, Dearborn, MI.)
vision use must be able to withstand continuous operation with repetition rates
of 30 times a second (10).

Fiber optics is the light guide that allows the transmission
of radiant power
through fibers (thin solid tubes) made of materials such as glass, fused silica, or
plastic (13). Most applications of fiber optics until today have been in the areas
of telecommunications and computer networking. However,with technological
45
Computer Vision
L i g h t Soorcc
W orkpiece
(b)
46
Panlgrahi and Gunasekaran
W orkpiece
Light
Sources
(r)
Fig. 3 Continued
47
Computer Vision
r s
\I
Light
Source
Light
Source
Cross-Polarizer
Translucent Part
1Polarizer
Light
Source
(i)
(1)
Fig. 3 Continued
49
Computer Vision
Workpiece
(1)
ul
0
Table 1 Comparison of Different Illumination Systems

Illumination System
Diffuse front illumination
Directional front illumination
Light tent
Advantages
Soft, fairly nondirectional
Reduces glare on metallic surfaces
Relatively easy to implement
Easy to implement
Good for casting shadows
Fiber optic delivery in many configurations
Eliminates glare
Eliminates shadows
Collimated back lighting
Produces very sharp edges for accurate gauging
Dark field illumination
Illuminates defects
Provides a high contrast image in some applications
Easy to implement
Creates silhouette of part
Very high contrast image
Low cost
Diffuse backlighting
Disadvantages
Edges of parts may be fuzzy
Low contrast on monocolor parts
May create unwanted shadows
Illumination is uneven
Must surround workpiece
Can be costly
Size can be a problem
Difficult to implement if material handling interferes
May be too bright for camera without neutral density
filters
Does not illuminate flat smooth surfaces

Difficult to impIement if material handling interferes
Low angle illurnination
Shows topological defects
Polarized front illurnination

Polarized backlighting
Eliminates glare
Highlights certain types of features or defects
in translucent materials
Relatively easy to implement
Strobed illumination
Structured light
Coaxial lighting
Crisp image with no blurring

Can be area. fiber optic, or light emitting diode
(LED)
Very long lifetime
Shows 3-D information
Produces high contrast on most parts
Laser frequency can be easily band pass
filtered
Eliminates shadows
Uniform illumination across FOV
Single source will produce uneven lighting across

surface
Reduces amount of light into the lens significantly
Only works for birefringent features
Difficult to implement if material handling interferes
More costly than standard sources
Requires accurate timing with camera
Must be shielded from personnel
Lasers above 5mW pose safety issue
Hard to image on some metals and black rubber
Complicated to implement
Harsh illumination for shiny surfaces
Source: Courtesy of Machine Vision Association. Society of Manufacturing Engineers. Dearborn. Michigan.
52 Gunasekaran
and
Panigrahi
advancements, fiber optics has been used for sensors as well as computer vision
applications. For computer vision applications,fiber optics has been used mostly
of fiber optics
for illumination and image transmission. It is expected that the use
as part of an image illumination subsystem of the computer vision system will
dramatically increase in the future.
In many computer vision applications, becauseof space and environmental
or to
constraints, it is necessary to provide illumination from remote locations
transmit image information to a remotely located computer system. Under these
conditions, the use of fiber optics is justified ( 1 1).
Fiber optics is highly desirable under the following conditions ( 1 1 ) :
Space for positioning the light source is restrictive.
The application requires camera movement along with movement
of the
illumination.
Multiple lighting with varied angle of incidence is required.
Maintenance of temperature profile of heat sensitive objects is critical.
Insertion of light through a small opening is required.
The light source could be hazardous in an explosive environment.
The application requires examination at micro or macro levels.
Optical fibers used for light transmission, illumination, and image transmission can be of the high-loss type, made up of optical glass or plastics. High-loss
type fibers are also cost-effective as compared to low-loss type fibers, which are
used primarily for data transmission, communications networks and other sensing
applications (12,13).
Fiber optics operates on the principle of total internal reflection. Optical
fibers exploit this phenomenonby encasing a cylindricalfiber in another cylindrical casing of lower refractive index (12,124). The outer casing is called clad,
and the inner fiber is called core (Fig. 4).
When light enters the fiber end at an angle A (incident angle), it undergoes
refraction at the interface of the core and its surrounding. Generally, the core has
a higher index of refraction (n,) than that of clad (11~). Under this condition, if
the angle of incidence (A) is greater than or equal to A, (critical angle), the light
will undergo total internal reflection (multiple times) at the core-clad interface.
Finally, it will leave the fiber through the outer end ( 1 19).
Using Snellslaw for optical fiber, its numerical aperture(NA) can be determined from the refractive indices of the core (n,) and cladding (n?), respectively
( I 19):
NA
[(nf - n;)]
NA
n sin (A,)
53
Computer Vision
Fig. 4 Illustration of totalinternalreflectance

(Adapted with permission from Ref. 124.)
of incidentlight
inan
opticalfiber.
where n is the refractive index of the surrounding medium. Since for air n = I
(for many cases, air is usedassurroundingenvironment),
N A = sin A,. N A
represents the ability of the fiber to accept light. Thus, the higher the NA, the
greater is the amount of light entering into the fiber. Therefore, to allow more
light through thefiber, consideration should be givento choose larger fiber (core)
and N A ( 1 19).
For both illumination and transmission applications, several criteria need to
be considered before selecting the appropriate optical fiber for a given computer
imaging system ( 120).
Range of operating wcrvelength: Most computer vision applications deal

with the visible spectrum (380-700 nm). However, with the growth of
it is important to know the
the nonvisible computer imaging systems,
range of the operating wavelength.
Numerical aperture: The required numerical aperture of the optical fiber
to be used.
Muterid ofopticcI1,fiher: For illumination applications, high-quality optical
glass is used for the fibers light transmitting core (13). For image transmission, glass fiber or other typesof low-loss fiber can also be used ( 12).
For nonvisible imaging applications, the right type of materials need to
be chosen such that the material can transmit electromagnetic energy
54
Chalcogenide
Fluoride
Glass
Enhanced
Glass
Silica
I
0.1
0.2
0.3
Ultraviolet
I
0.5
0.7
Visible
I l l
1
I
2
3
Wavelength
- (wm)
..
.
IIIII
7
10
20
Near-infrared
Fig. 5 Opticalfibermaterialsandtheirspectraltransmission(Adaptedwithpermission
fromRef. 12.)
within the operating wavelength range. Figure 5 shows different optical

fiber materials and their spectral transmission characteristics (12).
Overcrll length and trnnsnlissiorl loss: Though the type of application will
determine the overall length of optical fiber, it is important to note that
a transmission loss of energy (attenuation) occurs, which is dependent
on the overall length of
fiber used. Therefore,it is recommended to obtain
the transmission loss of optical fibers for different lengths and to optimize the appropriate overall length of optical fiber ( 1 19).
Activejber diatneter o r j b e r hurldle diameter: In computer vision applications, fiber bundles are used most often. A fiber bundle consists of several single
(bare) strands of optical fiber thatarearrangedparallel
to each other and are
used mostly for image transmission. Noncoherent fiber bundles imply a random
arrangementconsisting of optical fibers andareusedmostlyforillumination
applications (12,13). Nevertheless, the overall active diameter of a fiber bundle
is important because it affects the acceptance capability of optical fiber and is
also related to the numerical aperture (12,13,1 19,124).
Skeatlzirlg incrterial: This refers to the tubing or material that protects the
fiber bundle (13). There are different types of sheathing: steel. poly (vinyl chloride) (PVC), aluminum, and others. It can be flexible, rigid, or semi-rigid (120).
Sheathing on optical fiber greatlyaffects its performanceanddurability. The
Computer Vision
55
selection of sheathing depends on the operational environment, cable length, and

bending radius of the fiber ( 1 3 ) .
Minirnurn herding radius: Optical fibers are flexible. However,they should
not be bent more than their recommended minimum bending radius, which can
cause breakage in the fiber or even increase losses (12,13,124). It is advisable
to get the minimum and optimum bending radii of optical fiberdfiber bundles
from the manufacturer. A general rule of thumb is that bending a fiber more than
25 times its core or bundle diameter can damage the
fiber (13).
Environrnerlt: This refers to the environment in which the optical fibers
will operate (120). Temperature, liquid, solid, and other parameters (such as exposure to explosive gases, radiation, etc.) will affect the materialof fiber optics,
sheathing, and other manufacturing considerations ( 1 3,120). Plastic fibers can be
or explosive environments, a
used for applications below 80C. For hazardous
cold-light fiber optic system (eliminating the infrared component of light) can
be used ( 1 3 ) . Rigid fiber systems (fiber rods), where fibers are completely fused
to each other, are of particular value for measuring in and through liquids. An
ordinary light-conditioning glass fiber willturnbrownveryquicklyunderthe
influence of radioactivity. Typical rad-hard fibers are used that can withstand
long periods of radiation exposure ( 13).
Camera
C.
The camera is the sensor of a computer imaging system used to capture image
information. It functions similar to the eyes in human vision. Charged coupled
device (CCD) cameras have been used for nearly all computer imaging applications since their introduction almost 25 years ago (14). There have been many
developments basedonthe CCD technology.Recentlycomplementarymetaloxidesemiconductor(CMOS)technologyhasbeenintroduced(14,17).Many
varieties of black-and-white and color cameras are commercially available. A
black-and-white camera will suffice for food quality evaluationif attributes unrelated to color (dimensions, shape, other geometrical attributes) are to be evaluated. To evaluate color or color-related attributes, an appropriate color camera
should be selected. Color cameras range in cost from several hundred to several
thousand dollars. For higher-quality color images, three CCD color cameras perform better than one CCD color camera.
Though most cameras available are of area-array types, line-scan cameras
arealsoavailable in bothcolorandmonochromemodes.Area-arraycameras
are useful for imaging 2-D scenes, while line-scan cameras offer high positional
accuracy, rapid frame rate, and a wide dynamic range (14). They are available
in resolutions that include 128, 256, 5 12, 1024, 2048, 4096, and 8 196 pixels per
line. Many line-scan cameras have square pixels also (14).
For some food product evaluation, the nature of the inspection might re-
56
quire the camera to deal with low-light level images. This requirement can be
fulfilled by using another variationof line-scan camera, time-delay integration
(14). These cameras typically have 1024 lines and a number
of stages or rows
of sensors positioned side by side horizontally. Because these sensors also incorporate rows of photo elements, multiple views of objects can be captured. The
electrical charge from each rowis transferred from row to rowin synchrony with
the moving object to eliminate blurring. These cameras provide high signal-tonoise ratios as compared to their line-scan camera counterparts (14).
For imaging 2-D scenes, the most commonly used cameras are either frame
transfer or interline transfer CCD types. A frame transfer type camera has both
a sensing array and a storage array.
The storage array is positioned above the
sensing array. At the end of a field period (1/60 s, RS-170 signal), the data are
rapidly shifted vertically from the sensing array into the storage array. After the
data have been shifted into the storage array, they are shiftedto a horizontal shift
register, two lines at a time. The advantage of the frame transfer device is that
is
the entire area of the sensing array is sensitive to light. Their disadvantage
that the transfer process takes longer
and, thus, under certain conditions where
objects move, error in image data may be introduced (1 5).
Interline CCD type chips are used for most CCD cameras in the market.
of their low cost and their
They are attractive for many applications because
ability to handle bright localized
light overloads without streaking (15). They
can image moving objects without blurring, provided the scene has sufficient
light (1 5 ) .
Recently, progressive scan technology based on
CCD chips has created
progressive scan cameras that can be used for imaging fast-moving objects (14).
The emergence of these cameras has increased the potential of integrating computer-imaging techniques for high-speed, on-line applications in a cost-effective
manner. Progressive scan means noninterlaced or sequential line-by-line scanning
of the image information out of CCD as opposed to traditional interlaced fields
(16). Although this technology has been available for some time, especially for
frame-transfer type cameras, the recent introduction
of cameras basedon interline
transfer-based progressive scan devices has created new application advantages,
especially for high-speed imaging (16).
In these cameras, technology depends
heavily on effective shuttering (16). The technological advantages of interline
progressive scan ( 1 6) now can be applied to all types of high-speed and precision
imaging applications related to food quality evaluation.
Cameras for computer imaging that use charge injection devices (CIDs)
are also available. These sensors differ significantly from typical
CCD sensors.
Although the sensing pixels are constructed of metal oxide, semi-conductor capacitor integrating sites, the collected photon charge is read out differently ( I 5).
Major advantagesof CIDs includerandom access, radiation tolerance, nonbloom-
Computer Vision
57
ing, and adaptive exposure control (14). One disadvantage

is that the read-out
noise is higher than that from the conventional CCDs (15).
New innovations are observedin the camera market based on advancements
in CMOS technology. This has made possible the development of a single chip
camera that could also be very cost-effective (17). For food quality evaluation
where cost-effectiveness is much desired, it promises new applications.
A new image sensing technique involving active pixel sensors manufactured on CMOS shows additional potential (14). In the design of an active pixel
sensor, both the photo-detector and read-out amplifiers are integrated
at the pixel
site (14). These CMOS active pixel sensors are very new, and their integration
needs to be assessed for different food quality evaluations. A survey
of different
types of cameras based on CMOS chips
is given in Ref. 18.
Improvements in CCD chips have also created additional opportunities for
food applications. Thinned, back-illuminated CCDs (BCCDs) overcome the performance limitsof conventional front-illuminated CCDs by illuminating and collecting charge through the back surface (19). Cameras made up of BCCD show
higher overall sensitivity than that of a traditional CCD, which makes themvery
useful for fluorescent imaging and near infrared (NIR) imagingor imaging at the
far end of the visible spectrum (700-1000 nm). Use of BCCD has allowed the
capture of an image without using intensified CCD (19).
As discussed earlier, it is important to match the application requirements
with a cameras capabilities when selectingan appropriate camera. The following
parameters are also critical when selecting a suitable camera:
Resolution of the camera
Signal-to-noise ratio
Signal output
Minimum illumination required
Analog or digital output
Additional camera adjustment capabilities
D.
FrameGrabber
The basic function of a frame grabber or digitizer is to capture an image from

a camera so that a computer can access the image. Generally, a frame grabber
takes an analog video wave, samples it at specific intervals, and translates the
as an array of picture
information into a digital form (image), which is stored
elements (pixels) (43).
The selection of a frame grabber for a given computer vision application
is very critical. The following describes the critical features to be considered in
selecting a frame grabber.
58 Gunasekaran
1.
and
Panigrahi
Video Input
It is critical to ensure that the output of an image source (analog camera, digital
camera, VCR, etc.) matches the input of the frame grabber. The camera can be
monochrome (black-and-white) or color providing standard outputs as RS-170
monochrome or National Television System Committee (NTSC) color signals
(60 Hz video signals are used throughout North America and Japan) ( 1 18). The
camera can also provide Comite Consultatif International Radio (CCIR, an international organization) monochrome or phase alteration line color signals (50 Hz
or progressive
video standard)( 1 18). Nowadays, other cameras, such as line-scan
scan, are also available with nonstandard video outputs.
It is also possible to take image information from a video cassette recorder
(VCR) or digital camera.
It is often thought that digital cameras
do not need frame
grabbers. However, most frame grabbers need
an interface with a computer. It
is recommended to verify the proper required interface of the digital camera.
or of low
Some video signals, including those from VCRs, can be noisy
quality, resulting in blurry image acquisition due to missing or extraneous sync
(synchronization)signals.Thus,specialtimingcircuitry
is requiredonframe
( 1 18).
grabbers to correct for missing, extraneous, or low-level sync pulses
It is sometimes necessary to acquire monochrome images from color signals. The color (chrominance) content of these signals can cause interference
patterns, which degrade the quality of the image. Thus, frame grabbers are provided with a hardware/software selectable chrominancefilter that produces better
images ( 1 18).
2. Analog/DigitalCapability
Important aspects of analog/digital (A/D) capability arebriefly described below:

CI.
Spatial Resolution.
umns ( 1 18).
The number of pixels represented as rows
col-
h. BrigktrwssResolution.
Thisrepresentsthemaximumdiscrimination
of a given
capability of digital pixel value representing the brightness or color
pixel. It is expressed by the resolution of the AID converter, which is expressed
in number of bits (1 18). For a resolution of 12 bits, the resulting number of gray
levels is 2.
For color images, three A/D converters simultaneously convert red, green,
and blue information. For color frame grabbers, brightness resolution
is determined by the sum of all the A/D converters resolutions, i.e., 3 X 11 bits. For
example, a color frame grabber with
an 8-bit A/D converter has a brightness
resolution of 3 X 8 or 24 bits. The total number of colors is 2 or 16,777,2 I6
(25,43,118).
Computer Vision
59
c. Speed of A / D Cotwersiotl. The speed of A/D conversion in a frame

grabber is expressed in megahertz. It is very important that the speed of A/D
conversion is such that the required spatial resolutionof the image can be salnplcd
by the frame grabber.
For example, a 5 12 column X 480 row RS- I70 image signal needs 52.59
ps for each line. Each rowof the image of 5 12 columns requires 5 I2 A/D conversions. The time required will be ( 1 18):
52.59 ps/S 12 = 103 ns = 9.74 MHz
I O MHz
cl. Squrrre Pixels. Thegeometry of theimageshouldbesuch

that the
number o f pixels represents equal distance both horizontally and vertically. This
is called a square pixel (20). Regardless of the image shape or resolution (201,
it shouldbeverifiedfromthemanufacturer
that the framegrabbergenerates
square pixels.
e. G r q Scale Noise. In the framegrabbercircuit,randomness of noise
can cause variations in the digitization process (20). One way to define precision
in the digitization process is by evaluating gray scale noise (20).
Itis usually
represented by the least significant bit (LSB), a binary number representing the
gray scale value (20).
For example, an 8-bit frame grabber uses an 8-bit binary number to represent each gray scale. A change in LSB of % 1 implies a change of 2 1 gray scale
unit. A precision frame grabber typically has a gray scale noise of 0.7 LSB (20).
Gray scale noise can be a problem for biological and food applications when
dealing with low-contrast images (20). Applications such as analysis
of defects
(with subtle gray scale variations from adjacent good regions) can be affected
by the gray scale noise (20).
3. SignalHandling/Conditioning
The quality of an incoming video signal directly affects the qualityof a digitized
image. Poor lighting conditions, signal loss due to long cables, irregular sync
of camera are some parametersthat can cause
signals, and gray scale nonlinearity
poor quality in an incoming video signal. Well-designed frame grabbers can compensate for many of these problems by having different in-built capabilities (20).
Gain control adjustments, offset adjustments, and sync timing are three very desirable capabilities for frame grabbers (20).
(1.
Guin Adjrrsftnents. Gain adjustments can help in many situations such
as when adjustment of illumination is not possible and when the application depends on natural lighting (20). Other applications which are inherently of low
contrast (typically food/biological applications) or those which deal with passive
60
infrared imaging can also benefit from this feature. Thus, a gain control
frame grabber provides more versatility (20).
in the
b.OnsetAdjustments.
Offsetadjustment is verycritical,especially
if
the camera does not compensate for low or high lighting conditions. For best
results, a frame grabber should have the capability to offset a video signal by
2 100% in small, precise increments (20).
C.
ProgrammableGainand
Offset Controls. Becausegainandoffset
adjustments are very complementary, it is desirable that a frame grabber have
the capability for programmable control for gain and offset (20,118). The frame
grabber that is to be used with resettable cameras should detect each horizontal
sync pulse and instantly resynchronize its pixel timing. (Resettable cameras are
mostly used for industrial applications such as inspecting parts on conveyer
belts.)
For such applications, a desirable frame grabber is one with proper sync-timing
capability such as a crystal-controlled digital clock that can resynchronize instantly (20).
To digitize video
Another parameter to be considered is pixel jittering.
images, frame grabbers sample analog video at uniform intervals on each line to
determine the gray scale levelof each pixel (20). However, the inability
of frame
grabbers to precisely adjust the sampling points relative
to horizontal sync causes
pixel jitter (20). Phase locked loop (PLL) is the traditional timing mechanism
used for sync timing. PLL creates a clock from reference frequency.Pixel jitter,
in other words, is the timing accuracy of this clock and is expressed in nanoseconds. Pixel jitter of PLL varies greatly dependingon design and implementation
(21). For example, a frame grabber with 2 5 ns implies a pixel positional error
of 12.5% (20). Though frame grabbers are available that use both digital clock
circulating and modified PLL for precision applications with pixel jitter of *2
ns, pixel jitter of ? 10 ns is reasonable. The selection of acceptable pixel jitter
depends on the application (21).
4. SystemArchitecture for Integration

Another important set of criteria is the integration of the frame grabber with the
overall system. Criteria include the interfacing bus, in-board memory, video output support and digital I/O support (20).
a. Inteflucirzg Bus. Thoughframegrabbersinitiallyweredesigned
to
be interfaced with computers using industry standard architecture (ISA), present
technology allows using a higher speed bus. A peripheral component interconnect
(PCI) bus is an integral part of todays IBM-compatible, high-performance personal computer system.A PC1 bus has a theoretical bandwidth of 132 megabytes
per second and a continuous band width of 70-80 megabytes per second. It can
accommodate 32- and 64-bit data sizes(121). Although a PC1 bus has greater data
handling and transfer capability than provided by an ISA/EISA bus(1 18,12 11, it
of a given
is important to evaluate the data output and handling requirements
Computer Vision
61
imaging application. Many frame grabber manufacturers offerPC1 bus mastering

capability with frame grabbers. In this situation, the frame grabbers make the
required requests for memory transfers, freeing the CPU for other processing
tasks (22). It is also important to obtain additional information from the manufacto handle data. Some manufacturer on the adopted hardware desigdarchitecture
turers provide an on-board memory/frame buffer (with PC1 bus), while others
provide first-in, first-out (FIFO) memory (23). There are, of course, cheaper versions of frame grabbers with no memory at all. The selection of one type over
another depends on the application (22,23).
6. Digital Input/Output. Forrealworldapplications,

it isoftennecessary to correctly coordinate the timing of the image capture. One of the digital
input/output (I/O) capabilities of the frame grabber would consist of
a single
output strobe and a single input trigger (20).
The ability to generate a programmable width pulse is also desirable because camera control becomes more convenient with it (20).
a single frame grabIf the application requires using multiple cameras with
ber, the cameras are synchronized in a process called genlocking so that the
frame grabber does not encounter different video timing (20). Although many
cameras provide genlocking, variations and nonstandardizationof required camera input makes the realization of the process difficult. On the other hand, it is
very convenient to genlock the cameras from the frame grabber. Thus, for such
genlocking applications, the camera should supply horizontal and vertical drive
outputs (20).
Another desirable digital I/O capability of a frame grabber is delayed
trigger. This capability is very helpful for real-time industrial applications and
ensures that the interfaced camera will always capture images without failure
(20).
5. MiscellaneousCharacteristics
The following isa list of characteristics to considerin addition to those described
already when selecting a frame grabber (20).
a 12 V output for supplying power

Power output: some frame grabbers offer
to a camera, thus eliminating the need for ora separate power supply to
the camera (20)
Many frame grabbers have in-built digital signal processors (DSP) or dedicated hardware to perform different image processing operations
Support for operating systems and high-level programming languages such
as C/C+ + and Pascal Visual Basic (20)
Support for third-party programs/libraries for image processing (20)
Technical support/warranty
62
IV.
IMAGE PROCESSING
A.
Basic Steps
The basic steps in image processing are image preprocessing, segmentation, and
feature extraction (Fig. 2). The purpose of image preprocessing or image conditioning is to enhance the quality of the acquired image, which is often degraded
by distortion and noise in the optical and electronic systems of the input device.
(24): noise reducImage preprocessing steps include one or more of the following
tion, geometrical correction, gray level correction, and correction of defocusing.
These steps are typically applied uniformly and are context independent.
As thenameimplies,imagesegmentationrefers
to theprocessofsegorobjects.
menting or partitioningacompositeimageintocomponentparts
Proper segmentation is very critical. Often, the first step in assuring successful
segmentation is control of background uniformity. For monochrome images, segmentation normally is performed by examining the gray scale histogran-a bar
chart of the number of pixels in the image at different gray levels. Segmentation
algorithms are based on discontinuity or similarity of the gray level values. Discontinuities in image gray scale indicate sharp changesin image brightness such
as background and object.
In general, autonomous segmentation is one of the most difficult tasks in
image processing (25). Macaire and Postaire (26) described a real-time adaptive
thresholding to be used for on-line evaluation with line-scan cameras.
Segmented image data constitute raw pixel data of the image boundary or
or region
a region of interest in the image. The image representation as boundary
should be selected based on the intended application. For example, boundary
representation is appropriate for image size and shape characterization.
The region representation is suitable for evaluating image texture and defects.
The feature extraction step is thekey in deciphering the require image data
form the composite image information.The successof the feature extraction step
of the previous steps, including image
depends largely on the appropriateness
acquisition. The knowledge of the feature under consideration is also critical
at this stage in designing appropriate algorithms to extract information pertaining
to the desired feature(s).
Featureextractionfacilitatesobtainingsomequantitativeinformation
of
interest, which is then processed in conjunction with the knowledge base available for the feature studied.
6 . KnowledgeBase
At all steps during image processing, interaction with the knowledge base enables
more precise decision making. Thus, knowledge about the system being studied
of an image-processingsystem.Withoutan
shouldbeanintegralcomponent
Computer Vision
63
appropriate knowledge base, the vision system cannot think and make intelligent decisions (27). This problemis further complicated by the fact that the output
of a vision sensor is a complex combination of many parameters: size, shape,
texture, color, etc. Some requirements for intelligent decision making are (a) the
ability to extract pertinent information from a background of irrelevant details,
(b) the ability to learn from examples and generalize this knowledge and applyit
in different circumstances, and (c) the ability to make inferences from incomplete
information.
of
Expertsystems,neuralnetworks,andfuzzylogicaresomemethods
building knowledge bases into computer memories, enabling them to recognize
and interpret image data and to provide on-line control capabilities. The image
understanding part of the computer vision systemis inherently tied with the completeness and accuracy of the valid knowledge base available for the product(s)
and the feature(s) being studied. The successful image understanding step will
lead to the ultimate goal-translating image analysis data into information useful
for further action such as process/machine control. Applying neural networks
and/or fuzzy logic in conjunction with computer vision systems is rapidly growfor quality sorting of fruits and vegetaing, and commercial systems are available
bles (28).
C. PatternRecognition
Pattern recognition at some level is fundamental to image analysis. A pattern is
ina quantitative or structural description of an object or some other entity of
is formed by one or more descriptors
terest in an image. In general, a pattern
(features). Pattern recognition by machine involves techniques for assigning patterns to their respective classes automatically and with as little human intervention as possible.
In machine recognition of image patterns and shapes, generally two approaches are used: a statistical or decision-theory approach, in which features are
extracted and subjectto statistical analysis, and a syntacticor structural approach,
in which image primitives are selected and subjected
to syntax analysis.
The statistical or decision-theory approach is the traditional approach to
1960s. The system (Fig. 6)
pattern recognition that has been studied since the
consists of two parts: analysis and recognition. In the analysis part, a setof image
features that are judged to be nonoverlapping (or as widely apart as possible) in
the feature space is chosen (29). A statistical classifier (e.g., based on a fuzzy
logic or neural network system) is designed and trained with the chosen set
of
features to obtain the appropriate classifier parameters. In the recognition part,
an unknown image is filtered or enhanced in the preprocessing stage, followed
by feature detection and classification. This approach, however, does not describe
or represent structural informationin a pattern that is often desirable or necessary
64
Gunasekaran Panigrahi and
Classification
t """_
RECOGNITION
-------------"""""
ANALYSIS
t "_
I
Sample Pattern
I
Input
Image Pattern
rL
Decomposition
,.
FPrimitive
& Relation
Recognition
RECOGNITION
"--""""""""_
ANALYSIS
"1
Training
Selection
Syntax or
Structural Analysis
'
"
"
"
~
"
"
"
"
"
"
"
"
"
"
~
il
Relation Selection
Grammatical or
Structural lnferena3
b,
(8)
Fig. 6 Pattern recognitionsystems: (A) statistical and (B) syntactic. (Adaptedfrom
Ref. 29.)
for certain applications, as, for example, when the number of so

classes
large is
or
the given pattern isvery complex. In these circumstances, the numberof features
required is probably very large, making the statistical approach impractical.
In the syntactic or structural approach, complex patterns are decomposed
so on, until meaningful
into subpatterns and recursively into sub-subpatterns and
primitive patterns (analogous to features in the statistical approach) can be reliably extracted from them (29) (Fig. 6B). This approach allows us to describe
and represent the input pattern, in addition to classifying it into a specific class.
This approach has attracted much attention in the recent development
of pattern
recognition research.
D. Image Morphology
Image morphology refers to the geometric structure within an image, which inA general
cludes size, shape, particle distribution, and texture characteristics.
Computer Vision
65
approach in analyzing image morphology is to transform the given image to another where the information represented in the transformed image is more easily
understood. For example, investigationsof the shape of objectsin a binary image
a minimal set of pixels
often use thinning algorithms. Reducing an object to
representing an invariant of the objects geometrical shape is called thinning. A
skeleton is a line-thinned caricature of the binary image that summarizes the
shape and conveys information about its size, orientation, and connectivity (25).
An image resulting from the thinning process has many fewer black pixels representing the object and is, therefore, easier to manipulate.
If themaingoal of
thinning is data reduction and exact reconstruction of the original image is not
essential, many techniques are available that yield acceptable skeleton representations. However, if close or exact reconstruction is desired, care must be taken in
choosing an appropriate algorithm.
Morphological image-processing algorithms (thinning,
region filling, thickening, pruning,etc.) remain a useful tool in image processing and computer vision.
Some of the requirements for image thinning are (30):
Connected image regions must thin to connected line structures.
Approximate end-line locations should be maintained.
Thinning result should approximate the medial lines.
Extraneous spurs caused by thinning should be minimized.
The morphological approach has been successfully applied to a wide variety of problems. The power and usefulness of some basic morphological processing algorithms have been illustrated by McDonald and Chen (31). Morphologicalprocessingforisolated,nontouchingobjectsiseasilydoneusing
commercialpackages, which canperformobjectcounting
and dimensional
measurements, etc. However, touching and overlapping objects pose problems
unique tothe products being examined. Thus. separate algorithms and procedures
need to be developed. McDonaldand Chen (31) developeda morphological algorithm to separate connected muscle tissues in an image of beef ribeyes.
Recently, Ni and Gunasekaran ( 3 2 ) used imagethinning in conjunction
with a syntactic approach to evaluate the morphology and integrity of touching
and overlapping cheese shreds (Fig. 7). The algorithm performed very well with
less than 10% error in individual shred length measurements.
Evaluation of an image skeleton was also used to characterize granular
foods that may agglomerate (33). Smolarz et al. (34) used morphological image
processing to define structural elements
of extruded biscuits and then to discriminate biscuit type.
E. ShapeFeatureExtraction
The statistical or decision-theory approach has been widely used for food shape
it isoften
featureextraction.Foodmaterialshapeisveryimportantbecause
66
Gunasekaran Panigrahl and
Fig. 7 Morphological image processing for evaluating integrity of cheese shreds: (A)
cheese shreds, (B) binarized image, (C) after image thinning. (Adapted from Ref. 32.)
closely related to quality. Due to the demands of high quality, automated food
shape inspection has become an important need for the food industry. Due to
large inhomogeneities of food materials, however, such invariant shape features
cannot be used to detect local defects.
In many cases, therefore, the invariant
feature extraction methods cannot accurately distinguish between damaged and
undamaged categories. Panigrahi et al. (35) evaluated invariant moments and
fractal geometry for shape classificationof corn. Recently, variant shape extraction methods (position, orientation and scale) are gaining popularity for food
material shape inspection (36).
In the variant method, the edge contour of the inspected object is transformed to a given position, orientation, and scale. Then
the shape features are
extracted from every local edge point. Ding et
al. (37) presented a statistical
model-based variant feature extraction method for shape inspection
of corn kernels. This was based on a reference shape, a transformed average shape
of some
undamaged corn kernels. After the reference shape was obtained, the shape
of
kernels being inspected was compared with the reference shape.
(38) proposed a new algorithm with
More recently, Ding and Gunasekaran
improved ability to adjust object location, orientation, and scale to determine the
edge contourof a numberof food materials for shape evaluation. This multi-index
active model-based feature extractor is based on a reference shape comparison
Computer Vision
67
principle. The basic idea is to first transform and adjust of

a set
undamaged training objects to a certain location, orientation, and scale to obtain the average
of
transformed good object edge contour known as the reference shape. The second step is to transform and adjust each object to the same location, orientation,
and scale. Then the shape of the objects under inspection can be compared with
the reference shape to identify any significant deviations. Corn kernel and animal
crackershapeinspectionisusedasexamplefoodmaterialstoillustratethis
method. The reference shape contour of an undamaged product is indicated by
the dotted linein Fig. 8. The line going through the geometrical center, the origin,
and another (prespecified) point could be considered as the x-axis. Then a number
of equal-angular locations are chosen, starting from the zero angle direction (xaxis). An arbitrary equal-angular location (ern]) and the corresponding radius
(R[k]), the distance from origin to kernel edge, are shown on Fig. 8. A number
of shape indices pertaining to object radius, curvature, continuity, symmetry etc.
can be calculated and used for identifying damaged objects. The multi-index
approach resulted in a more accurate identificationof damaged objects than the
single-index approaches used in the past.
F. Image Texture
Texture is characterized by the spatial distribution of gray levels in a neighborhood. For most image-processing purposes, texture is defined as repeating patterns of local variations in image intensity, which are too fine to be distinguished
(30). Thus, a connected setof pixels
as separate objects at the observed resolution
Fig. 8 Food shape evaluationto detectdamaged products by comparingobject and reference edge contours: (A) corn kernel, (B) animal cracker. Object edge contour has been
converted into Transformed edge contour and comparedwith Average of transformed
good kernel edge contours. (Adapted from Refs. 27 and 38.)
68 Gunasekaran
and
Panigrahi
satisfying a given gray-level property that occurs repeatedly in an image region

constitutes a textured region. A simple example is a repeated pattern of dots on
a white background. Image texture can be usedto describe such image properties
as smoothness, coarseness, and regularity (25). There are three approaches to
studying image texture characteristics: statistical, structural, and spectral.
Statistical methods are used extensively
in texture classification, identifying
the given textured region from a given set of textured classes. Image data such
as mean, standard deviation, and moment (a measureof the frequency of occurrence of pixels of a given gray level within a particular image region) are used
to study smoothness, coarseness, graininess, etc. Techniques are also available
to study additional image texture characteristics such as entropy (randomness)
and uniformity.
Structural techniques of image texture analysis deal with the arrangement
of image primitives such as the descriptionof texture based on regularly updated
parallel lines. Spectral methods are based on the Fourier transform to study the
global periodicity of an image. For example, presence of high frequenciesin the
frequency image may represent a coarse texture. Gonzalez and Woods (25) have
further described the structural and spectral methods in some detail.
Image texture analysis can be performed using either monochrome or color
(39) usedgray-scaleimagecharacteristics
to studybread
imagedata.Zayas
crumb grain properties. Tan et al. (40) used HSI space image texture properties
to evaluate corn extrudate characteristics. Ruan et al. (41) performed texture analysis on RGB image data for evaluating wheat kernel features.
V.
COLORIMAGE PROCESSING
Color is an important property of biological and food products. Color variations

play a major role in quality and disease evaluation. Modern computer imaging
systems are capableof acquiring and processing color images. Although the discipline of computer imaginghision is nearly 40 years old, the color computer imaging technique is relatively young (10-12 years old).
The basic differences between a gray level and a color computer imaging
system arein the camera, frame grabber, and display
monitor-the camera should
of handling
be a color camera, the frame grabbeddigitizer should be capable
color information, and the display monitor should be a color monitor capable of
displaying color information.
A.
ColorCoordinates
Color computer imaging systems represent color information in terms of color

coordinates. Several types of color coordinates are found in color theory. How-
Computer Vision
69
ever, RGB and HSI havebeen extensively used for color image processing applications. Each color coordinate has three components. The combination of these
components produces a color image.
1. RGB
In the RGB color coordinate system, the three color components are the three
primary or fundamental colors red, green, and blue. Different combinations
of
these primary colors produce various secondary colors. The Commission Internationale de 1Eclairage (CIE) uses the spectral primary system RGB to define any
color by combining red, green, and blue (light generated by a monochromatic
light source at 700, 435.8, and 546.1 nm, respectively) (43). The chomaticities
r, g, and b (normalized red, green, and blue) are defined as:
r=R/R+G+B
g=G/R+G+B
b=B/R+G+B
The RGB color coordinate system is commonly used in television, color
cameras, and color display monitors. In the television industry, RGB signals are
encoded into luminance (Y) and chrominance (I and Q) to minimize bandwidth
for facilitating broadcast (43). RGB cameras and monitors have
been used for
computer graphics because RGB isa good system for generating and displaying
images. Similarly, digitizers use three A/D converters to digitize the RGB signal.
The digitized color imageis stored in three color components/buffers, which are
mixed together only during display to show one composite color image.
According to Travis (42), RGB technology is fine for grabbing, storing,
in RGB space is computationand displaying the image, but processing the image
ally intensive and algorithmic implementation is complex. Moreover, RGB is a
poor way to represent images basedon human vision because people do
not think
of color in terms of combinations of red, green, and blue.For example, it would
be difficult to look at a yellow cake and specify the percentages of red, green,
and blue that combine to form the color of the cake.
2. HSI
The HSI color coordinate system is an alternative to the RGB system. Hue (H)
is defined as the attribute of color perception by means of which an object is
judged to be red, yellow, green, or any other color. Intensity (I) represents the
attribute of color perception by means of which an object is judged to reflect
more or less light than another object. Saturation (S) represents the attribute of
70
color perception that expresses the degree of departure from gray of the same
lightness.
Understanding and manipulating color images in HSI is much easier and
less complicated than in RGB because the HSI system resembles the way we
perceive color. Individual values of H, S, and I contain information that is meaningful for human color perception and can be analyzed independently. Thus, the
algorithmdevelopment is lesscomplicated(43,122).Thefollowingempirical
relationships between RGB and HSI color coordinates make it possible to switch
between the two (25):
I =
S
R + G + B
3
= 1
(6)
3 .
-[mm(R,G,B)]
I
H = cos
[(R
(7)
[(R - G) + (R - B)]/2
G) + (R - B)(G - B)].5
where if B > G, then H = 2n: - H.

Note that much digitizer/frame grabber and commercial image processing/
analysis software is able to convert color images from one set of coordinates to
another. It is possible that some modifications might be incorporatedin the equations above for their hardware implementation. Some manufacturers even use
different existing relationships to convert from RGB to HSI and vice versa. It is
recommended that these empirical relationships be obtained from the manufacturers.
In addition to RGB and HSI, other color coordinates have been used for
color image processing applications. Details can be found in other publications
(43,44).
B. Considerations for Color Imaging Systems

Appropriate color camera and digitizers are definitely critical for color imaging
systems. As emphasized earlier, selection of the appropriate light source (illumination source) is critical. In addition to several other considerations, when selecting a gray level-based imaging system, two more factors need
to be taken into
account when selecting light sources, especially for the color imaging system:
color rendering index and color temperature/chromaticity of the light source.
Color rendering involves the property of light to reveal an objects colors (45). The color rendering index (CRI)
of a light source measures the degree
of color shift that objects undergo when illuminated by the light source as compared to the color shift of the same object when illuminated by another light
Computer Vision
71
source ( I 23). On a CRI scaleof 0- 100, 100 represents a source with the renderof 0 (zero) implies an illumination source
ing capabilities of daylight (45). A CRI
incapable of rendering color. Thus, the higher the CRI, the more vibrant or brilliant the color is. Light sources with a CRI of 80 or higher have excellent color
rendering properties. A CRI of 70-80 implies a good color rendering property
(45,123).
Color temperature is the absolute temperature of a black body radiator
having a chromaticity equal to that of the light source (123). It is expressed in
of 3000 K (low) corresponds to warm or
degrees Kelvin. A color temperature
red-yellow appearances. Light sources at 3500 K provide neutral white light and
those at 4100 K provide cool bluish light. Light sources witha color temperature
of 5000 K give off daylight (45).
Both CRI and the color temperature of a light source can be obtained from
the manufacturer. For developing a color computer imaging system, i t is always
recommended to select a light source with CRI above 85 and a color temperature
close to 5000 K (123). Another secondary but important parameter
is the stability
of the color temperature of the light source over time. The color temperature of
many light sources changes with time. Thus, it is recommended to select a light
source whose color temperature does not change with time.
C.ColorCalibration
Calibration of a color computer imaging system impliesthat all the critical comto
ponents (i.e., camera, frame grabber, display monitor) should be calibrated
handle, process, or display color information. Thus, calibration of a color computer imaging system includes the calibration of its components, such as a color
camera, frame grabber, and display monitor.
Many end users or developers, unfortunately, have not practiced color calibration. Often it is assumed that all components are working satisfactorily. In
many cases, however, a small deviation
of calibration of one component can
introduce errors in the final result provided by the color computer imaging. For
example, if a color camera looking at an orange puts out the color information
as yellow or red instead of orange, then error is introduced.
1. DisplayMonitorCalibration
Video test generators or test pattern generators available commercially can be
The generatorsendsdifferentsingle-color
connected to thedisplaymonitor.
charts such as red, green, blue, etc. and white, black, or multiple color charts to
the monitor to be displayed on thefull screen. The user thencan adjust the monitor for proper calibration for a wide range of color conditions or for user-desired
specific color condition. At present, computer add-in boards are also available
72
that can be used in place of stand-alone video test or pattern generators. A list
of manufacturers of video test generators are listed in Ref. 46.
2. CameraCalibration
A standard method to calibrate a cameras output uses two pieces of test equipment: an NTSC vectroscope and a waveform monitor. Vectroscopes and waveform monitors are extensively used by the television industry, and they complement each other. Waveform monitors display the video signal to allow the user
to measure its amplitude and time parameters. At the same time, the vectroscope
can show the relationship between a chrominance signal and its reference burst
(or phase) and gain distortion (46).
or
Under a given lighting condition, the camera can look at standard, single,
multiple color bar charts. The camera output is sent simultaneously to a waveform
monitor and an NTSC vectroscope. The waveform monitor measures the amplitude of a video signal, and the vectroscope shows the relationship between color
information (46). Necessary adjustments canthen be madein the cameras control
unit to calibrate the camera.
Note that an NTSC vectroscopeonly accepts NTSC composite input,
which
most analog cameras provide. If the camera does not have an NTSC output,
individualRGBoutputscanbeinterfacedwithRGBinputs
in any calibrated
video display monitor. In this case, the NTSC vectroscope is not used. Using
visual observation of the displayed color on a calibrated video monitor, proper
adjustments can be made in the color camera.
3. FrameGrabberCalibration
After the camera and display monitor are calibrated, the frame grabber can be
calibrated. Standard single-color bar charts typically havea 94-96% reflectance
value. Images can be acquired separately using single red, green, blue, white,
and black charts. For red, green, and blue conditions, the average pixel values
of the respective buffer for the digitized images should be about 252-255. Any
deviation can be addressed by changing the gain and offsets of respective A/D
converters of the frame grabber. Note that the frame grabber has three A/D converters for each of the red, green, and blue channels. The frame grabber should
have programmable gain and offset adjustment capabilities. Similarly, the frame
grabber can be adjusted under white (presence of all color) and black (absence
of all color) conditions. Different researchers have related incorporationof calibration of different color imaging components or systems. Panigrahi (47) described calibration of a color imaging system for corn quality evaluation, and
Hetzroni and Miles (48) described color calibration of RGB video images. Several researchers have reported additional procedural
and mathematical techniques
for calibrating cameras, even in on-line conditions (49-51).
Computer Vision
73
D. ColorImageProcessingApplications
In recent years, color computer imaging technology has been extensively applied
to numerous food-related applications, which can be broadly grouped into color
evaluation, defect detection, and texture evaluation
of different food products,
including dairy, meat,fish, fruit, vegetables, and others(52). This variety of applications, however, presents challenging color image processing issues. These applications can be discussed under two broad image processing categories: color
segmentation and image analysis and understanding.
1. Color Segmentation
In processing color images, segmentation refers to isolating or separating a homogenous or desirable region of interest in the image. Removal of a background
from an image is a simple example. Segmentation is also used to identify and
quantify all sorts of defects, diseases, and other abnormalities. The segmentation
problem in color image processing is more complex than in gray scale applications. A color imageis comparable to three gray level images having color information contained in three color components, e.g., red, green, and blue. Thus,
segmentation becomes more time-consuming and involved for color images. Of
course, the level of complexity depends significantly on a given application. The
inherent random variability of quality attributes of raw materials (agricultural
products) for food products further adds to the complexity
of segmentation of
color images of many food products.
Thresholdingbasedonhistogram
is usedforapplicationsthatseparate
background from the object or separate two
or three dissimilar contrasting regions
in the image. One requirement is that there should be a good amount
of color
difference among the regionsto be segmented. Sometimes investigationis neces(54).
sary to choose the appropriate color coordinate for performing segmentation
To evaluate the color
of French fries, for example, removing background information from the French fries was required. Finding the threshold based on histogram
on the RGB color component was difficult and time-consuming. Using the HSI
coordinate, a histogram was obtained on the I (intensity) component. Finding a
threshold in the intensity image alone was easier than with the RGB image. Using
the determined threshold, the background was isolated from the image. All backgroundpixelswerelabeled
in intensityimageandsubsequentlymappedinto
saturation and hue images. Thus, for color evaluation, the hue and saturation
information of background pixels were not considered (54).
Use of adaptive thresholding techniques for a histogram-based segmentation is also recommended for food images. They promise higher accuracy and
robustness than a fixed (global) threshold. Adaptive thresholding techniques can
adapt to changes in lighting and spectral characteristics of an object as well as
the background. Therefore, they are well suited for real-world applications and
74 Gunasekaran
and
Panigrahi
most food quality evaluation applications. The description of different adaptive

thresholdingtechniquesincludingothertraditionalimagesegmentationtechniques such as region growing, clustering, and region merging can be found
in
Refs. 43, 44, 54, and 55.
With the recent advent
of neural network technology and
its associated
advantage of being fault-tolerant intelligent, it has been used for unsupervised
segmentation of color images (56-58). Unsupervised neural networks are best
suited for real-world images. They do not need supervision or a teacher. as
do
supervised neural networks, in order to conduct segmentation. The self-organizing map (SOM) has been extensively studied for unsupervised image segmentation. This type of neural network can work for multidimensional data such as a
color image having three-dimensional color information.It preserves the images
topography and simultaneously will maintain its spatial relationships. Details of
the architecture of SOM neural networks can be found in Ref. 58. Though applications of SOM networks or other neural networks for food image segmentation
have not been reported extensively in the literature, the success of their applications on natural color images(59,60) and other multidimensional pattern recogniof neural network
tion techniques (6 1 ) clearly reinforces the potential success
technologies for image segmentation of food products.
2. ImageAnalysisandUnderstanding
In analyzing color images of food, the exact techniques for image analysis and
understanding differ from applicationto application. For example, an image analof corn might
ysis and understanding algorithm developed for color classification
not work fully with high accuracy for potatoes or potato products. This provides
additional challenges and requires investigation for developing successful applications.
Selecting appropriate color coordinates for analyzing a given food color
image is critical. An accurate answer to the question, Which color coordinate
do I choose? can be obtained by experimentation only for a given application.
For color evaluation of edible beans, Panigrahi et al. (63) evaluated both r-g-b
(normalized RGB) as well as hue h-s-i (normalized HSI) coordinates. Both sets
of coordinates provided accuracy up to 100% in classifying beans in three color
groups (63). To identify and quantify fat in meat images, RGB color space was
usedwitharectangularprismand
Mahalano bois distance criteria (64). RGB
color coordinates were used for locating citrus fruits for harvesting (65) and RG
color space wasutilized along with Bayes decision theory for image partitioning
and subsequent color classification of stone fruits (66).
Recently, the potential of artificial intelligence technologies, such as neural
networks and fuzzy logic, has also been explored for image classification and
understanding purposes. Both neural network and fuzzy logic techniques are intelligent, adaptive, and fault-tolerant(57),and they complement each other. Neu-
Computer Vision
75
ral networks are a new paradigm of computing or information processing inspired

by biologicalmodels.Forcolorclassification
of Frenchfries,Panigrahiand
Marsh (54) condensed bothhueandsaturationhistograminformationbytwo
separate back-propagation neural networks.The condensed color information was
thenfedasinput
to another neuralnetwork,RProp,avariationofbackpropagation neural network. The maximum color classification accuracy obtained
by this modular network was 96% for classifying a given French fry sample into
three color groups (54). (In this case, neural networks were used
in a modular
format.) Similarly, another probabilistic
neural network was used for color classification of French fry samples into three color groups: medium, light, and dark
(67). A few multistructure neural network classifiers were used to classify four
of accuracy of 95.9%
varieties of pistachio nuts with an average classification
(68). Detection of blood spot and dirt staining on eggs was performed with an
accuracy of 85.6% and 80%, respectively, using neural networks (69).
These are only a few successful applications of neural networks for color
image classification; their applications are growing rapidly. A few potential neural networkarchitecturesareback-propagation,learningvectorquantization,
radial-basis neural network, and recurrent neural network. Details about these
neural network architectures can be found in Haykins (59).
Similarly, fuzzy logic is another intelligent information processing mathematical paradigm for dealing with uncertainty, vagueness, and ambiguity. Fuzzy
logic has been successfully used for real-world complex image classification and
understanding (70). It was used to diagnose tomato disease (71) and to analyze
and classify other biological and agricultural images (75). Extending
its use to
classify and evaluate food images is definitely very encouraging.
Rough sets theory, similar to fuzzy logic technology, has also been used
for defect detection and quality evaluation of edible beans based on their color.
A knowledgeThe maximum classification accuracy achieved was 99.6% (72).
of corn kernel propbased discrimination function was adopted for dissemination
erties along with a learning vector quantization network resulting in a classification accuracy of 95-100% for yellow and white corn kernels (73).
The complementary characteristicsof neural networks and fuzzy logic have
created a new technique called neuro-fuzzy system. Neuro-fuzzy techniques
in soybean seed with a maximum accuracy
have been used to classify disease
of 95% (74). Other applications of neural networks and fuzzy logic for image
segmentation and classification can be found in Ref. (75).
VI.
THREE-DIMENSIONALCOMPUTERIMAGING
Most of the food materials are three-dimensional (3-D) objects, and hence for a
complete and thorough evaluation of food quality, 3-D information is necessary.
Therefore, there is an increasing need for extracting 3-D information, which will
76
increase the capabilities of computer imaging systems similar to that of human

vision.
Though there are similarities between the 2-D and 3-D computer
vision
applications, there are also some differences in how the image is acquired and
represented. 3-D imaging is an extension of 2-D imaging where an additional
measurement depth or range provides the third dimension to the image. Figure 9 outlines different techniques that can be used to measure 3-D information
about an object (77). These methods are broadly grouped as direct and indirect
methods. In direct methods, the depthor range measurement is obtained directly.
However, in indirect methods, 3-D measurements are obtained indirectly from
2-D image(s) (77).
A.
Overview of Direct 3-D Measurement Techniques
1.
Time of Fhght
A time-of-flightrangesensor as describedbyNitzan (77) includes a signal

transmitterandsignalreceiver.Thesignaltransmittersendsthesignal
to the
target object. The receiver consists of a collector that collects a part of the signal
reflected by the target and other required electronics for measuring the round trip
travel time of the returning signal. The two types of signal generally used are
Fig. 9 Differenttechniquesformeasuring3-Dinformation.(Adaptedwithpermission
fromRef. 77. 0 1988 IEEE.)
Computer Vision
77
ultrasound and laser light. Beacuse

it is an active range sensor, the principle
of Lambertion reflections governs its operations. Therefore, it may not function
properly if the surface of the object being viewed at is highly specular (77).
One type of ultrasonic range camera is made by Polaroid (76), which uses
53, and 50 kHz. Range finders
ultrasonic signals at four frequencies: 60, 57,
based on ultrasonic systems are generally
not capable of being used for mediumto high-resolution applications. Still, for other applications, such as navigations,
this system can be used to determine the locations of impediments (76). Generally, ultrasonic range sensors have relatively low resolution because of their inability in properlyfocusingacoustichltrasonicsignalascomparedtoalaser
beam. Nevertheless, for applications requiring
low resolution, they might be more
cost effective than a laser-based range sensor (76,77).
According to Nitzan (77), time-of-flight laser range sensors generally use
a scanning mirror to direct the transmitted laser beam to sweep across the target
object with equal angular increments to obtain a dense range data consisting of
N X M range elements called rangels. There are two schemes for measuring
the time of flight between the object and the target in a time of flight laser range
sensor: pulsed time delay and phase shift. Figure I O depicts the configurations
of both of these. In a pulsed time-delay system,a pulsed laser is utilized to deter-
Target Pomt
I\
Beam
Transmitted
Transmitter
Laser
* Pulsed Time Delay
* Modulated (Frequency
Amplitude)
or
Ltght
\Recewed
Light
Recewer
Range
Intensity
Scannmg Mirror Unit

Reference Beam
Fig. 10 Configuration for pulsed time delay and phase shift laser range sensing systems.
(Adapted with permission from Ref. 77. 0 IEEE 1988.)
78 Gunasekaran
and
Panigrahi
mine the range based on the measured time of flight of the transmitted signal
(76,77). In a phase shift system, a frequency or amplitude modulated laser sends
the signal. The measured phase shift between the transmitted and the received
signal determines the range (77).
Laser-based range sensors could provide better resolution than ultrasonic
range sensors, but they are relatively costly (76,77). Moreover, slow measurements and ambiguity (when phase shift is greater than 360) issues create additional problems for laser range sensors (77). Nevertheless, the rapid reductionin
the cost of the laser systems, along with their increased capabilities, could eliminate some of these problems.
2.
Triangulation
Triangulation based on elementary geometry (Fig. 1 1 ) is defined by Nitzan (77)

as: Given the base lineof a triangle, i.e., the distance between twoof its vertices
and the angles at these vertices, the range from one of the vertices to the third
is computed as the corresponding triangle side. A simple range-finding geometry as described by Jarvis (76) is presented in Fig. 12. According to Nitzan (77)
Triangulation techniques are subdivided into two schemes: structured light us-
Epipolar Plane
Projector
(StructuredLtght)
Fig. 11 Illustrationofthetriangulationprocess.(Adaptedwithpermission
77. 0 1988 IEEE.)
from Ref.
Computer Vision
79
Baseline
Fig. 12 A simple range-finding geometry. (Adaptedwithpermission

0 1983 IEEE.)
from Ref. 76.
ing a projector of controlled light and a camera/detector(an active scheme) and

stereo using ambient light and two cameras (a passive scheme). The accuracy
of depth measurement that can be obtained using this technique depends on the
measurement accuracies of distances, positions and angles (79,117). Detailed
analysis of triangulation systems is given by Case et al. (78). The mathematical
concepts and relationships for using triangulation techniques to approximate 3-D
data can be found in Faugeras (79).
a. Srrucrured Lighting. Structured lighting refers to the process of controlling or structuring the light to form different patterns or structures on the
used. The first and most basic
object (78,81,83,84). Three common techniques are
technique allows the light to fall on the object in the form of a spot. Then, an
imaginary triangle is formed using light source, detector, and the point
on the
object (Fig. 13A) (78). This technique only allows the measurement of a single
pointlspot for each scan. Thus, to generate a 2-D image of resolution
pxq. a total
number of (pxq) scan would be required (78).
In the second technique, a light source is used to generate a single stipe
of light which falls on the object and the image is acquired by a2-D detector or
camera (Fig. 13B) (78,80,81). This arrangement allows the acquirement
of an
image using a fewer scans (compared to that required
by the first technique) (78).
80 Gunasekaran
and
Panigrahi
Object
Detector
(A)
(B)
(C)
Fig. 13 Different structured lighting systems:(A) simple system,(B) light stripe system,
(C) multiple stripe system. (Adapted with permission from Ref. 78. 0 Kluwer Academic
publishers.)
The width of light stripe defines the spatial resolution (84). If the object is moving, the vision system will needa few or no moving parts. However, the direction
of travel should be perpendicular to the light stripe (78).
The third technique uses multiple stripesof light simultaneously (Fig. 13C)
(78,82) and is very similar to the second arrangement (Fig. 13B). One disadvantage with this approach is that multiple stripes can cause ambiguity during the
process of image recognition. However, the use of several gray-coded light stripes
might solve this problem (82,83). The gray-coded light stripe method provides
computational advantages too. For example, if N ordinary light strips (N scans)
were required to completely scan an object, only log, N gray-coded light stripes
(log? N scans) would cover the entire object (81-83).
Bayer and Kak(1 16) reported the integration of color with structured lighting. This method of color-encoded structured lighting shows a few benefits, i.e.
increased speed and better accuracy. In this technique, they ( 1 16) used a single
encoded grid of colored light stripes to obtain range information. Grid to grid
alignment problems generally found in multiple stripe technique, were overcome
with this method (1 16). However, problems were encountered in dealing with
objects having deeply saturated colors. Thus, the applicability of this technique
might be limited to applications where the objects are of neutral color (1 16).
Many benefits are associated with structured lighting. Its inherent simplicity
to acquire 3-D information.
has madeit one of the most commonly used technique
it. Objects with high
Nevertheless, several difficulties are still associated with
specular surface characteristics could provide incorrect
or sometime very little
range information (76,77,82,83). Though structured lighting technique sometimes
could be slow in acquiring image information (77,82,83), recent developments
in high speed processors and detectors might eliminate this problem.
In cases
where triangulation techniques are used with structured lighting, sometimes hidin acquiring quality
den surface or edges
of the object might cause problems
images or processing acquired images (77,8 1,83,84).
Computer Vision
81
Fig. 14 A typical stereo configuration for 3-D imaging. (Adapted with permission from
Ref. 117. 0 SME)
Structured lighting has been used for several agricultural applications, some
of which dealt with production and animal agriculture. Structured lighting with
two-constraint propagation was used to measure 3-D surface featureson potatoes
(85). Structured lighting was also used to find stalk and calyx of apples during
high-speed detection of blemishes on apples (86).
Another 3-D measurement system using structured lighting was developed
to determine the shape of soybean seed, its axial length, surface area, volume,
particle density, compactness, and sphericity by Sakai and Yonekawa (87). In
at the center
their study. A soybean sample was mounted on a needle located
of a supporting table. A camera was placed over the sampleat a known distance.
to the
A vertical plane of light struck the sample at an oblique angle relative
cameras horizontal axis. A helium-neon laser light source was used, which emitted structured light vertically deflected by a polygon mirror at 10,000 rpm and
was narrowed by a biconvex lens. The systems performance was satisfactory
(87). However, more work is necessary to further develop the system.
b. Stereo Stereo is a popular 3-D imaging technique ( 1 17). This method

is also called stereo triangulation technique (77,117).In this method, an imaginary triangle is formed with the objectas one of its vertex. The other two vertices
of the triangle are generally meant for two cameras and are called camera vertices. Two static cameras can
be used on the camera vertices
to produce a
stereo pair of images. Sometimes one camera can be used but it must be transported between the two vertices of camera locations. A typical stereo system is
illustrated in Fig. 14 ( 1 17).
82 Gunasekaran
and
Panigrahi
Camera modeling and calibration are important procedures for 3-D information acquisition using stereo (79,117). According to Bachnak ( 1 17), Camera
calibration is the procedure for determining intrinsic and extrinsic parameters of
the camera. Such parameters include the focal length of the camera, the scaling
factor of the system, the transformation relationship between the 3-D scene or
object, and the image plane. Detailed description of camera calibration is mentioned by Faugeras (79).
Extraction of depthor3-Dinformationusingstereoinvolvesdisparity,
cameraparameters,andimagecorrespondence(76,88,117).Imagecorrespondence implies matching corresponding points in the stereo pair images (76). For
determiningcorrespondencefromimages,Jarvis(76)emphasizedthatthere
must be sufficient visual informationat the matching points to establish a unique
pairing. Two basic problems arise with this process.
The first occurs at parts
where uniformity of intensity or color makes matching impossible. The second
happens when the image of some part of the scene or object appearsin only one
view of the stereo pair because of occlusion effects or limited field of view captured in the images.
According to Bachnak (1 17), two regularly used approaches to matching
or correspondence are area-based and feature-based matching. Area-based approaches result in reasonably accurate disparity maps, but they are sensitive to
changes in contrast and depth. Feature-based methods, on the other hand, focus
on easily distinguished properties i n the images such as lines, corners, etc. The
result is normally an accurate sparse disparity map. Further discussion on tackling correspondence problems are mentioned by Yakimovosky and Cunningham
(88).
Another critical step to be emphasized for using the stereo techniqueis the
optimization of baseline or distance between two cameras (76,77,88,117) (Fig.
1 I). If the baseline is smaller than optimum, accuracy in 3-D measurement could
be affected. The largerthe distance than optimum between cameras, on the other
hand, could increase the accuracy of depth measurement. However, the problem
of hiddedmissing surfaces could occur (76,77,117). A stereo system developed
at NASAs jet propulsion laboratory for guiding a robotic vehicle
is described
by Matthies and Anderson (89).
B. 3-D Microscopy
Study of microstructure is one of the most fundamental ways of evaluating food
quality because the macroscopic textural properties are, in fact, a manifestation
of the microstructural arrangement of constituents of a complex food material.
Stanley and Tung (90)
defined microstructure as a complex organizationof chemical componentsundertheinfluence
of externalandinternalphysicalforces.
Foods having similar structures can be loosely grouped together as foods that
Computer Vision
83
have similar textures (90). Therefore, study

of microstructure has been well established. However, past studies on food microstructure are mostly based on subjective, qualitative assessmentof 2-D micrographs. Such subjective evaluations cannot provide enough information to quantify the effects
or establish interactions
amongvariousparameters.Therefore,imageanalysis
is oftenused to obtain
quantitative information from micrographs (91).
In conjunction with image analysis, microscopy can also be used for adaptive control of food fermentation and
other biotechnology applications (92).
Almost all instruments used to provide an enlarged view of food systems
a light or
can be used with image analysis, either through direct interface with
electron microscope or by scanning outputs such as photographs or negatives.
Image analysis can also help to understand mechanisms
of complex processes that
alter product characteristics. There are several ways to study the microstructure
of
food materials. The method chosen depends on factors such as the nature of the
food, the microscopic information of interest, and the level
of resolution required
(92). Electron microscopy offers the advantage
of high resolution, but sample
preparation procedures such as sectioning, dehydration, and chemicalfixation are
laborious and may lead to artifacts.
Confocal laser scanning microscopy (CLSM) offers an alternative way
to
observe food structure with high resolution but without disturbing the internal
structure. It is a powerful tool to penetrate a samples surface and to visualize
thin optical sections. These thin optical sections can be used to study the layered
2-D microstructure, and a computer algorithm can also reassemble them into 3D images for 3-D image analysis of their microstructure.
The basic principle exploited by the confocal microscope
is that of defocus (93). When a conventionally imaged object is displaced from best focus,
the image contrast decreases, but the spatially averaged intensity remains the
same. In a confocal imaging system, however, the image of a defocused surface
appears darker than if it were in focus. Thus, confocal optics can be said to
of
haveaxialresolution
in additionto lateral resolution.Asaconsequence
this property, it is possible to extract topographic information from a set of conof focal planes. In order to form a confocal
focal images taken over a range
image, the signal is recorded as the object is scanned relative
to the image of
the point source in a plane parallel to the focal plane. Multiple confocal image
slices are obtained by repeating the process at various levels of object defocus.
By focusing at different heights (along the z-axis) on the object, a 3-D topographical map of the object is obtained. The resolution in the z-direction (axial
resolution Az) depends upon the numerical aperture (NA) of the lens, the degree
to which the pinhole is open, and the wavelength of laser light. If the confocal
pinhole is fully opened, the microscope becomes a conventional scanning light
microscope with reduced lateral resolution and a larger depth of field in the zdirection.
84 Gunasekaran
and
Panigrahi
Oneprincipaladvantage
of opticalsectioningoverserialsectioning
is
avoiding physical specimen distortion dueto cutting and having image alignment
from the various image planes. Another advantage is depth resolution; while inferior to the lateral resolution in each plane by a factor of about 2-3,
it is still
useful for many applications (9). The minimum separation between observation
planes is 0.05 pm (94), a difficult resolution to achieve using regular light microscopy. Maximum observation depths of 10-100 pm can be achieved, depending
on a specimens opacity and absorption characteristics.
The CLSM has been a proven technique for a number
of biomedical applications. However, it is still in its infancy for food quality evaluation (3). Its application for studying food materials is expected to increase within next few years
due to the ability of CLSM to:
Penetrate deeply but noninvasively into the specimen
Obtain large numbers of sequential, thin optical sections that may be reassembled by a computer to produce 3-D images or stereo pairs
to threeorfourseparatechemicalcomponents
Identifyandlocalizeup
(depending on the number of laser lines available on the instrument) by
using specific fluorochrome labeling techniques (95)
In addition, specimen observations can be made within a plane both transverse to and along the optical axis, as compared with conventional light microscopy, which can only make images transverse to the optical axis. Sample preparation for CLSM involves staining the lipid or aqueous
protein phase of cheese with
fluorescent dyes and subsequent observation after laser excitation (97). However,
CLSM is limited by the maximum possible magnification, which
is about 400.
Vodovotz et al. (96) provide details of CLSM functioning.
Brooker (97) presented some figures of mozzarella cheese microstructure
obtained using CLSM.Hassan et al.(98) used CLSM to observe coagulum formation resulting from milk acidification. Vodovotz et al. (96) and Blonk and van
( 1 00)
Alast (90) reviewed many other CLSM applications. Ding and Gunasekaran
have developed a 3-D image reconstruction procedure to build a 3-D network of
fat globules in cheese. The 3-D view of fat globules in Cheddar cheese is presented in Fig. 15. Throughthisreconstruction,informationabouttheglobule
volume and related properties were measured and related to cheese-making procedures, which were not otherwise possible (101).
Fig. 15 3-D microstructureevaluationof fat globulesinCheddarcheese.(A.

B. C)
Selccted 2-D layered images of full-fat, low-fat. and very-low-fat Cheddar cheeses, respectively (the width of each microscopic image is
77 pn). (D) Reconstructed 3-D view of
fat globules in low-fat Cheddar cheese. (Adapted from Ref.
101.)
Computer Vision
85
Layer 16
86 Gunasekaran
and
Panigrahi
CLSM not only allows visualization of in situ 3-D microstructure, it also

allows quantification of some important features. Because 3-D analysis provides
comprehensive data from a sample much larger than those used for typical 2-D
microscopy, 3-D data are more accurate and less dependent on sample location
and viewing direction. However, the limitation of observation depth
is a potential
problem. Quantificationof 3-D image features obtainable with CLSM could serve
as objective criteria for evaluating quality or the effectof a number of variables
of interest in cheese making.
Data-handling requirements of 3-D images are very extensive. For exam13megabytes of storage
ple, 50 layers of a 512 X 512imagerequireabout
space. Image processing isalso computationally intensive, and, depending on the
computers CPU, it may take several days to completely process a single 3-D
image. Image storage requirements may be minimized by compressing individual
2-D slices, for example, by the joint photographics expert group
(JPEG) compression algorithm. There is still no standard image compression algorithm (9).
Ding and Gunasekaran (100) developed
a computer algorithm to reconstruct a 3-D network of fat globulesin cheese from sequential 2-D layered images
obtained from a CLSM. A few 2-D slices and the reconstructed 3-D image
of
fat globules in cheese are presented in Fig. 15. Thus, the 3-D image processing
technique has helped us, for the first time, to evaluate in situ 3-D characteristics
in
of a fat globule to understand the effect of process parameters and fat level
cheese textural qualities.
Due to the development of novel imageprocessingtechniquesandimof computers, 3-D microscopy is fast
proved computational power and speed
becoming the latest trend in microstructural analysis of foods.
VII.
NONVISIBLECOMPUTERIMAGING
Although the majorityof computer imaging technology uses the visible spectrum
(380-700 nm), the nonvisible electromagnetic spectrum also has potential for
(UV),
use in computer imaging. These nonvisible bands include x-ray, ultraviolet
near-infrared (NIR), and infrared (IR). Advances in semiconductor-based detector technology and declining component prices have triggered the integration of
nonvisible imaging techniques for food quality evaluation.
A.
FluorescentImaging
Most food products or raw materials of food products can use fluorescent imin the next chapter. Formost cases
aging. Principles of fluorescence are described
of fluorescent imaging, the wavelengths used range from the far end ofUV (300
Computer Vision
a7
nm) to the far end of VIS (700 nm). Intensified CCD cameras have been used
for this typeof application. Because of the low amount of signals available, these
intensified CCD cameraswork better than a conventionalCCD camera. Recently,
however, the introduction of DSPs with conventional CCD cameras has made it
possible to create low-light cameras for acquiring fluorescent images without
using intensified CCD. These low-light cameras have the ability
to vary the time
1/60 or 1 / 130 second to several minutes.
of integration of image information from
By integrating a weak fluorescent signal for a longer time, a quality fluorescent
image is obtained.
The introduction of BCCD has also generated another viable option for
acquiring quality fluorescent and NIR images.
The spectral sensitivityof a BCCD
camera is significantly higher than that of intensified
CCD and conventional CCD
cameras, especially at the UV and NIR ends of the visible spectrum (19).
B.
NIR Imaging
NIR images can be very valuable for food quality evaluation. For imaging pur700-1 100 nm and
poses, the NIR waveband can be divided into two groups:
> 1100 nm.
Because of the higher sensitivity of BCCD cameras in the lower NIR region, they can be used for NIR imaging of food products. Similarly, some monochrome CCD cameras have relatively high sensitivity in the lower NIR region.
Although the sensitivity of monochrome CCDs in the 900- 1 100 nm zone is not
as high as that of a BCCD, there is a big difference in cost. Thus, which camera
one chooses to use depends on the application. Note that before using a monochrome CCD camera for NIR imaging, the IR filter in front of the CCD sensor
head must be removed. It is also highly recommended that the sensitivity curve
of the camera be obtained from the manufacturer
to verify that the camera is
appropriate for the application.
NIR imaging can also be achieved by using a liquid crystal tunable
filter.
to a standard CCD detector to produce
Tunable filters can be easily coupled
digital images at any wavelength within 400- I 100 nm (102). It has no moving
parts. Since it is capable of acquiring images at many wavelengths,it can be used
to generate multispectral images (102). Note that the quality of the image still
depends on the sensitivity of the CCD detector used.
NIR images based on 700-1 100 nm can be used for detecting defects and
formappingmoisture (970 nm)andprotein (1020 nm) in foodproducts. An
example is detecting defects in peaches. A monochrome CCD camera witha
band pass filter centered at 750 nm (with a bandwidth of 40 nm) produced the
images. The images were further analyzed for placing a peach into one of eight
classes based on different defects. The classification error based on NIR images
was 3 1 % compared to 40% obtained with color images (103).
88 Gunasekaran
and
Panigrahi
The NIR spectrum (700-2500 nm)

is sensitive to chemical constituents,
e.g., protein, moisture, and oil of food and agricultural products. Although NIR
spectroscopic techniques have been used for quality evaluationof food products,
NIR imaging could provide additional spatial information
that is not available
from traditional spectroscopic signals. For example, NIR spectroscopy can be
used to measure the overall protein, oil, or moisture content, whereas NIR images
will show the distribution of such constituents within the food material. ThereIt is
fore, NIR imaging may replace NIR spectroscopy for some applications.
in conjunction with
more likely that NIR imaging/visible imaging may be used
visible/NIR spectroscopy. Park et al. (104) integrated multispectral imaging (using 542. 571. 641, 700, 726, 847 nm) with visible/NIR spectroscopy (417-965
nm) for inspection of poultry carcasses.
NIR images above 1 100 nm can be obtained using indium-gallium-arsenide
(1nGaAs)-based cameras available from Sensors Unlimited (Princeton, NJ). Area
to
cameras are sensitive to 900-1700 nm. and line-scan cameras are sensitive
800-2200 nm. Both cameras produce analog and digital output, including
RS170. They can be operated at room temperature, thus eliminating the need for
cooling (105). These capabilities show tremendous promise for integrating nonvisibleNIRtechnologyintoevaluationand
analysis of food composition and
constituents in a nondestructive manner. Most food constituents such as protein,
oil,water,starch,sucrose,glucose.andotherchemicalsbasedonhydrogencarbon bonds have been evaluatedby spectroscopic methods. NlR imaging would
provideadditional spatial informationthatspectroscopycannotprovide.With
these capabilities, functional or compositional images
of food products can be
acquired. which can help quality evaluation and inspection and also provide information on the interaction of food components that could be valuable for product
development and quality evaluation.
C.InfraredImaging
Focal plane array thermal infrared cameras without liquid nitrogen cooling (using
a sterling cycle-based cooler instead) are now available from commercial sources.
Some of them are compact and easy to use and provide better spatial resolution
and thermal sensitivity. They are sensitive to the thermal infrared band (3-5 p n )
and can capture images at 30 frames/s with 12-bit dynamic ranges. With emissivity and atmospheric correction capabilities, they can create thermal images
of
food products. 1R cameras can also measure temperatures from - I O to 1500C.
Thus, IR cameras promise another rapid and nondestructive technique for food
quality evaluation. especially for characterizing thermal properties, thermal mapping, and moisture-related studies. 1R imaging was used to estimate the internal
temperature of chicken meat after cooking (106).
Computer Vision
89
D. X-Ray Imaging
X-rays are another component
of the electromagnetic spectrum. They contain
high energy and can be used for nondestructive imaging. Recently, the development of filmless and low-energy x-ray detectors has created expanded possibilities for x-ray imaging for food and agricultural applications.
X-ray line-scan imaging was used to classify apples based on water core
features (107). X-ray technology was used to predict grain yield (108). Although
the integration of x-ray technology might find some obstacles for food quality
evaluation from consumers, low-energy x-ray devices might gain acceptance
in
the future for nondestructive evaluation where other imaging techniques will not
work.
VIII. ON-LINE OR MOVINGSCENEANALYSIS

Most computer vision systems designed in the past were concerned primarily
with static scenes. However, the perception of visual motion plays an important
role in many emerging computer vision applications. Thus, computer vision systems to analyze dynamic scenes are being designed.
Input to a dynamic or moving
a changing world.
scene analysis systemis a sequence of image frames taken from
in motion. Each
The camera used to acquire an image sequence may also be
frame represents an image of the scene at a particular instant in time. Changes
in a scene may be due to the motionof the camera, the motion of objects, illumination changes, or changes in an objects structure, size, or shape. It is usually
assumedthatchanges in a scene are due
to camera and/or object motion. A
system must detect changes, determine the motion characteristicsof the observer
and the objects, characterize the motion using high-level abstraction, recover the
structure of the objects, and recognize moving objects.
Depending on the designof the imaging system, different image processing
techniques are required. Recovering information from a mobile camera requires
a different technique than one suitable for images from a stationary camera (30).
In the food industry, the most common design is that of stationary camera and
moving objects. Image input is a frame sequence represented by F(x,y,t), where
x and y are the spatial coordinates in the frame representing the scene at time t.
The value of function F represents the intensity of the pixel.
In many applications, an entity, a feature or object, must be tracked over
a sequence of frames. If there is only one entity in the sequence, the problem is
easily solved. With many entities moving independently in a scene, tracking requires the use of constraints based on the nature of the objects and their motion.
A number of real-time visual tracking algorithms are described in Eklund et al.
90 Gunasekaran
and
Panigrahi
(109). Due to inertia, however, the motion

of a physical entity cannot change
instantaneously. If a frame sequence is acquired at a rate such that no dramatic
change takes place between two consecutive frames, then no abrupt change in
motion can be observed for most physical objects (30). This has been the basis
of most on-line applications currently available in the food industry. The important factor is then to set the image acquisition rate fast enough to minimize image
by frame. Real-time image
blur so the analysis of image data can take place frame
processing boards and real-time processors are available to assist in on-line realtime computer vision applications ( 1 10).
For a continuous stream of material flowing down a conveyor belt, a computer vision system can be designed using a line-scan camera for image acquisiof photosensitive sites.
tion. A line-scan camera contains a one-dimensional array
The line-scan camera is suitable for fairly fast moving object scenes.In addition
to higher speed, line-scan cameras offer high resolution and the ability to handle
infinitely long image scenes. A new breed of cameras, known as time delay and
CCD image sensor technolintegrate (TDI) cameras, are line-scan cameras using
ogy to gain an increase in speed or sensitivity of up to 100 times that of conven( 1 1 I ) . A 2-D image
tional cameras while providing exceptional spatial resolution
can be produced if there is relative motion between the camera and the object
of interest. The columns of information from the line-scan camera are usually
stored sequentially in a framestore allowing interpretation of returned data as a
2-D image. The authors research team is currently evaluating such
an on-line
system to evaluate the qualityof shredded cheese. The run-length coding (binarizing an image in which each pixel is a 1 or 0) of the binary image is used to
identify object locations in the scene (a string of Is represent an objects presence). Syntactic pattern recognition technique was used in the image interpretation step.
Use of strobe lighting is also an effective technique for acquiring on-line
information from a moving scene.To obtain a complete imageof the scene under
strobe lighting, the strobe firing must be synchronized with camera and image
acquisition. Lack of synchronization will appear as partially light and/or partially
or totally dark digitized images. The most straightforward strobe and image acquisition synchronization is where an object present is a signal typically generated
by a photo eye or a proximity switch device. In this technique, the strobe light
is fired immediately on an objects arrival, so the amount of object placement
uncertainty in the direction of travel is reduced significantly. However, this technique requires high-precision object sensors, a special television camera with the
capability of scan inhibit, and an electronic circuit that synchronizes the various
to avoid image blur during
timing signals. Ni et al. ( 1 12) used a strobe light
image acquisition to examine individual grain kernels. The kernels were traveling
at a modest speed of 6.5 m/min.
Computer Vision
91
General requirements for on-line applications are throughput (speed), accuracy, consistency, durability, diversification,
flexibility, and adaptability. Considerations of these conditions and constraints have to be given atall stages of
system design and development. Speedof evaluation is perhaps the most striking
requirement. For example, Tao et al. (8) estimated that an on-line apple grading
system may have to examine at least 3600 fruit/min. They describeda computer
vision system sorting 3.5 million fruit in an 8-hour day, which consisted of two
separate lighting unitswith eight camerasand one processing and control console
unit. This type of machine is being widely installed in the packing industry for
sorting applies, citrus, peaches, tomatoes,
and various other fruitsand vegetables.
IX. SUMMARYANDFUTURETRENDS
Computer vision techniquesplay a significant rolein fulfilling the needs of rapid
and nondestructive sensors in food quality evaluation. The increasing demand
for real-time or high-speed food quality evaluationwill be met by recent developments in high-speed DSP integrated cameras and frame grabbers. Commercially
available DSP chips (e.g., Texas Instruments C-40, C-44,
and C-80; Intels I860) offer flexibility, expandability, and upgradability suitable for parallel processing of images. Image processing analysis operation is a better candidate for
parallelism. Therefore, parallel processing can be integrated for high-speed and
real-time applications using the DSP chip.
At present, many commercial DSP
integrated frame grabbers are available that can be configured for parallel processing ( 1 13). Dedicated high-speed processors also offer alternate solutions for
high-speed and real-time imaging applications.
Recentdevelopments in busarchitecture,such as compact PC1 (CPCl),
high-speed networks for image transmission, and cost-effective storage devices
will add to the increasing integrationof real-time imaging systems for food quality evaluation ( 1 14).
The development of portable, cost-effective, miniature laser systems permits the implementation of structured lighting techniques for extracting 3-D informationforfoodproducts.Moreover,ongoingtechnologicaladvancements
have made 3-D cameras available from commercial sources. With the capabilities
of high-speed processors and cost-effective 3-D imaging software, food sectors
will encounter increased 3-D computer imaging for automated quality evaluation
and inspections.
Significantgrowth in thetechnologicalcapabilities of software has occurred along with their user-friendliness and adaptability. One does not need to
know high-level programming in order to developa specific application. Capabil-
92
ities in commercial software to programin English-like macro languages would

allow users to develop the application quickly and easily.
of artificial intelligence technolThere will also be a progressive integration
ogies such as neural networks, fuzzy logic, and genetic algorithms to harness
their advantages of fault-tolerance, intelligence, and accuracy for food quality
evaluation (1 15). The capabilities of implementing neural networks and fuzzy
logic algorithms on hardware chips would also expand the integration
of these
intelligence technologies for real-time or high-speed food applications.
Thus, computer vision technology bears a tremendous potential as a nondeIt is
structive technique for a variety
of food quality evaluation applications.
obvious that this technology will play an important role in making our food products safe and of high quality.
REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.
IO.
I I.
12
P Locht. Full color image analysis

as a tool for quality control and process development in the food industry. ASAE Paper
No. 97-3006, St. Joseph, MI, 1997.
BG Batchelor. Lighting and viewing techniques. In: BG Batchelor, DA Hill, DC
Hodgson, eds. Automated Visual Inspection. Bedford, UK: IFS Publications, Ltd.,
1985,pp103-179.
HESchroeder.Practicalilluminationconceptandtechniqueformachinevision
applications. SME Technical Paper No. MS84-397, 1984, pp 397-408.
M Coletta, K Harding. Picking the right lighting. MVA/SME Vision Techno1
Q
7(1):1-3,1990.
S Gunasekaran, TM Cooper, AG Berlage, P Krishnan. Image processing for stress
cracks in corn kernels. Trans ASAE 30( 1):266-271, 1987.
BL Upchurch, SA Throop. Effects of storage duration on detecting watercore in
apples using machine vision. Trans ASAE 37(2):483-486, 1996.
SH Mersch. Polarized light for machine vision applications. Proceedings of the
Third Annual Applied Machine Vision Conference, International Society for Optical Engineering, Bellingham, WA, 1984, pp 4/40-4/54.
Y Tao, L Chance, B Liu. Full scale fruit vision sorting system design-factors and
considerations. Proceedings of the Food Processing Automation Conference IV;
ASAE, St. Joseph, MI, 1995,pp14-22.
JC Russ. The Image Processing Handbook. Boca Raton, FL: CRC Press, 1992.
ANovini.Fundamentals of strobelightingformachinevision.Proceedingsof
Vision 87 Conference, Society of Manufacturing Engineers, Detroit, MI, 1987, pp
4113-4/25.
CG Clarkson. Illumination for machine vision. In: Illumination Solutions for Machine Vision and Microscopy. Lawrence, MA:Doh-Jenner Industries, Inc., 1996,
pp 3.1.
M Muehlman. Optical fibers for illumination. In: The Photonics Design and Applications Hand Book. MA: Laurin Publishing Company Inc., 1991.
Computer Vision
93
13. Fibers for lighttransmission.In:ThePhotonicsDesignandApplicationsHand
Book. MA: Laurin Publishing Company Inc., 1988.

for machine vision applications. Vision
14. A Wilson. How to select solid-state cameras
Systems Design. Pennwell Corp., Tulsa, Oklahoma (July):56-60, 1998.
15. P Mengers. Solid-state area imaging devices (a survey and comparison). Paultek
Note No. 2082, Nevada City, CA, Paultek Corp., 1994.
16. RHofmelster.CatchingtheactionwithprogressivescanCCDs.ThePhotonics
Spectra(July):132-136,1995.
17. P Suni. Advanced design creates single-chip vision systems. Laser Focus World.
Pennwell Corp., Tulsa, Oklahoma (April):73-83, 1997.
a chip:Devicesarehere.Advanced
18. BCarlson.CMOSsensor-basedcameraon
Imaging Technology, 1999, pp 46-49.
19. GM Williams, NM Bloure. How to capture low-light level images without intensitiers. Laser Focus World. Pennwell Corp., Tulsa, Oklahoma (Sept.): 1995.
20. G Marshall. Choosinga frame grabber for performance profitability. Technical Report. Imagenation C o p , 1997,pp1-12.
a brush up. Ad21. S Lathe. Ten things to consider when choosing framegrabbers:
vanced Imaging, (Jan.):22-25, 1996.
22. A Wilson.Choosing a PCI-based frame grabber for vision applications. Vision
Systems Design. Pennwell Corp., Tulsa, Oklahoma (Aug.):46-56, 1998.
23. A Wilson. Diverse architecture complicate image processing board-design choices.
Vision Systems Design. Pennwell Corp.. Tulsa, Oklahoma (Dec.):53-59, 1998.
24. Y Shirai. Three-Dimensional Computer Vision.New York: Springer-Verlag, 1987.
25. RC Gonzalez, RE Woods. Digital Image Processing. Reading, MA: Addison-Wesley Publishing Company, 1992.
26. L Macaire, JG Postaire. Proceedings of Computer Vision for Industry, SPIE, 1989,
pp14-25.
27. S Gunasekaran, K Ding. Using computer vision for food quality evaluation. Food
Techno1 48(6): I5 1 - 154, 1994.
28. LA Berardinis. Untangling the mystery of neural networks. Machine Design, 1992,
pp 55-59.
29. EK Wong. Syntactic image pattern recognition. In: ER Dougherty, ed. Digital Imaging Processing Methods. New York: Marcel Dekker, 1994, pp 167-195.
30. R Jain, R Kasturi, BG Schunick. Machine Vision. New York: McGraw-Hill, Inc.,
1995.
31. TP McDonald, YR Chen. Separating connected muscle tissues in images of beef
carcass ribeyes. Trans ASAE 33(6):2059-2065, 1990.
32. H Ni, S Gunasekaran. A computer vision method for determining quality of shredded cheese. Special issue of AI Applications in Biology and Agriculture, Artificial
IntelligenceReview12:27-37,1998.
33. F Ros, S Guillaume, G Rabated, F Sevila. Recognition of overlapping particles in
granular product images using statistics and neural networks. Food Control 6( I ) :
37-43,1995.
34. A Smolarz, E Van Hecke, JM Bouvier. Computerized image analysis and texture
of extruded biscuits. J Texture Stud 20:223-228, 1989.
35. S Panigrahi, MK Misra, S Wilson. Evaluations of fractal geometry and invariant
94
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
moments for shape classification of corn germplasm. Computers Electronics Agric

20: 1-20, 1998.
FS Lai, I Zayas, Y Pomeranz. Application of pattern recognition techniques in the
analysis of cereal grains. Cereal Chem 63(2):168-172, 1986.
K Ding,RV Morey, WF Wilcke, DJ Hansen. Corn quality evaluation with computer
vision. ASAE Paper No. 90-3532, St. Joseph, MI, 1990.
K Ding,S Gunasekaran. Shape feature extraction and classification of food material
using computer vision. Trans ASAE 37(5):1537-1545, 1994.
IY Zayas. Digital image texture analysis for bread crumb grain evaluation. Cereal
Foods World 38(10):771-776, 1993.
J Tan, X Gao, F Hsieh. Extrudate characterization by image processing. J Food
Sci 59(6):1247-1250, 1994.
R Ruan, JZ Xu, RK Jones. Rapid analysis of scabby wheat using machine vision
and neural networks. Proceedings of the Food Processing Automation Conference
IV, ASAE, St. Joseph, MI, 1995, pp 59-66.
M Travis. Image processing goes color. Computer Graphics World (June):99-101,
1988.
AK Jain. In: Fundamentals of Image Processing. Englewood Cliffs, NJ: PrenticeHall, 1989, pp 60-77.
R Hunter, R Harold. In: The Measurement of Appearance.
New York: John Wiley & Sons, 1987, pp 119-237.
LampSpecificationandApplicationGuide.TechnicalManual,PhilipsLighting
Company, Somerset, NJ, 1996.
A Wilson. Off-the-shelf test equipment qualities video systems. Vision Systems,
(Sept.):78-85,1998.
S Panigrahi. Color computer vision for characterization of corn germplasm. Ph.D.
dissertation, Iowa State University, Ames, IA, 1992.
A Hetzroni, GE Miles. Color calibration for RGB video images. ASAE Paper No.
94-3007, St. Joseph, MI, 1994.
E Marszalec, M Pietikainen. On-line color camera calibration. Proceedings of the
IEEE Conference on Pattern Recognition, Jerusalem, 1994, pp 232-237.
L Robert. Camera calibration without feature extraction. Proceedings of the IEEE
Conference on Pattern Recognition, Jerusalem, 1994, pp 704-749.
Y Tao. Closed loop search method for on-line automatic calibration of multi-camera
inspection systems. Trans ASAE 41(5):1549-1555, 1998.
S Domenico, W Gory. Machine vision and neural nets
in food processing and packaging-natural combinations. Proceedings of the FPAC
III Conference, Orlando, FL,
1994, pp 11-19.
SKFu,JKMu. A survey on image segmentation. Pattern Recognition
133-16,
1981.
S Panigrahi, R Marsh. Neural network techniques for color classification of French
fries. ASAE Paper No. 96-3004, St. Joseph, MI, 1996.
S Panigrahi, MK Misra, C Bern, S Marley. Background segmentation and dimensional measurement of corn germplasm. Trans ASAE 38(1):291-297, 1995.
Q Yang. An approach to apple surface feature detection by machine vision. Computers Electronics Agric 1 1 :249-264, 1994.
Computer Vision
95
57. Y Pao. In: Adaptive Pattern Recognition and Neural Networks. New York: Addison-Wesley Publishing Company, Inc., 1989, pp 223-236.
C+ +.
58. AS Pandya, RB Macy. In: Pattern Recognition with Neural Networks in
Boca Raton, FL: CRC Press, 1996, pp 23-72.
59. S Haykins. In: Neural Networks: A Comprehensive Foundation. Macmillan College
Publishing Company, Inc.. 1994, pp 394-434.
60. J Iivarinen, A Visa. Unsupervised image segmentation with the self-organizing map
and statistical method. Proceedings of the SPIE Conference on Intelligent Robots
and Computer Vision XVII, Bellingham, WA, Vol. 3522, 1998, pp 516-525.
61. TUchiyama,MAArbib.Colorimagesegmentationusingcompetitivelearning.
IEEE Trans Pattern Anal Machine Intelligence 16( 12): 1 197- 1206, 1994.
62. SC Newton, S Pemmaraju, S Mitra. Adaptive fuzzy leader clustering of complex
data sets in pattern recognition. IEEE Trans Neural Networks 3(5):794-800, 1992.
63. S Panigrahi. Advanced information technologies for objective quality sensing of
edible beans. Proceedingsof Sensoral98, International Workshop on Sensing Quality of Agricultural Products, Montpellier, France, 1998, pp 88-98.
64. X Gao, J Tan, D Gerrad. Image segmentation in 3-dimensional color space. ASAE
Paper No. 95-3607, St. Joseph, MI, 1995.
65. F Pla, F Juste, F Ferri, M Vicens. Color segmentation based on reflection model
to locate citrus fruits for robotic harvesting. Computers Electronics Agric 9:5370,1993.
66. N Singh, MJ Delwiche, RS Johnson. Image analysis methods for real-time color
grading of stone fruits. Computers Electronics Agric 9:71-84, 1993.
67. Y Chtioui, S Panigrahi, R Marsh. Conjugate gradient and approximate methods for
an optimalprobablisticneuralnetworkforfoodcolorclassification.OpticEng
37(1 l):l-9, 1998.
68. A Ghazanfari, J Irduyaraj, A Kausalik. Grading pistachio nuts using a neural network approach. Trans ASAE 39(6):2319-2324, 1996.
69. VC Patel, RW McClendon, JW Goodrum. Detection of blood spots and dirt stains
in eggs using computer vision and neural networks. Trans ASAE 12(2):253-258,
1996.
70. TL Huntsberger, C Rangarajan, SN Jaya Ramamurthy. Representation of uncertainty in computer vision using fuzzy sets. IEEE Trans Computers C-35(2): 145146,1986.
71. C Wu, S Hu, W WU, J Lu. Fuzzy related computer vision for diagnosing tomato
diseases. ASAE Paper No. 93-3033, St. Joseph, MI, 1993.
72. Y Chtioui, S Panigrahi, L Backer. Rough sets theoryas a pattern classification tool
for quality assessment of edible beans. Trans ASAE 42(4): 1 145-1 154, 1999.
73. K Liao, Z Li, JF Reid, MR Paulsen, B Ni. Knowledge-based color discrimination
of corn kernels. ASAE Paper No. 92-3579, St. Joseph, MI, 1992.
74. IS Ahmad, JF Reid, MR Paulsen. Neuro-fuzzy inference of soybean seed. ASAE
Paper No. 97-3041, St. Joseph, MI, 1997.
75. S Panigrahi. Neuro-fuzzy systems: applications and potential in biology and agriculture. AI Applications 12(1) & (2): 1-13, 1998.
76. J Jarvis. A perspective on range finding techniques for computer vision. IEEE Trans
PAM1 5(2): 122-139 (0 1983 IEEE).
96 Gunasekaran
and
Panigrahi
77. D Nitzan. Three dimensional vision structure for robot applications. IEEE Trans
PAM1 10(3):291-309 (01988IEEE).
78. S Case, JA Jalkio, RC Kim. Three-dimensional vision system analysis and design.
In:Three-DimensionalMachineVision.Boston:KluwerAcademicPublishers,
1988.
79. Faugeras. In: Three-Dimensional Computer Vision:
A Geometric View Point. Cambridge, MA: MIT Press, 1996, pp 9-108.
80. GJ Agin, TO Prinford. Computer description of curved objects. IEEE Trans ComputersC-25:439-449,1977.
81. MJohannesson.Sheet-of-lightRangeImaging.ThesisNo.404,Departmentof
Electrical Engineering, Linkoping University, Linkoping, Sweden, 1993.
82. JL Mundy, GB Porter. A three-dimensional sensor based on structured light. In:
Three-Dimensional Machine Vision. Boston: Kluwer Academic Publishers, 1988.
83. P Besi, R Jain. Three dimensional object recognition. Computing Surveys 17( I ) :
75-145,1985.
84. M Altschuler, K Bae. Robot visionby encoded light beams. In: Three-dimensional
Machine Vision. Boston: Kluwer Academic Publishers, 1988.
85. H Gongzhu, G Stockman. 3-D surface solution using structured light and constraint
propagation. IEEE Trans 2(4):390-401, 1989.
86. Q Yang. Finding stalk and calyx of apples using structured lighting. Computers
Electronics Agric 8:3 1-42, 1993.
87. N Sakai, S Yonekawa. Three-dimensional image analysis of the shape of soybean
seed. J Food Eng 22 1-234, I99 1.
88. Y Yakimovsky, R Cunningham. A system for extracting three-dimensional measurements from a stereo pair of TV cameras. Computer Graphics Image Proc 17:
95-210,1993.
89. L Matthies, C Anderson. Stereoscopic vision system for robotic vehicle. NASA
TechnicalBriefs17(11):72,1993.
90. M Kalab, P Allan-Wojtas, SS Miller. Microscopy and other imaging techniques in
food structure analysis. Trends Food Sci Technol 6(6): 177-186, 1995.
91. S Inoue, KR Spring. In: Video Microscopy-The Fundamentals. New York: Plenum,1997,pp119-157.
92. JB Litchfield, JF Reid, SJ Schmidt. Machine vision microscopy and magnetic resonance microscopy. Food Technol 48(6): 163- 166, 1994.
a fast
93. AR Rao, N Ramesh, FY Wu, JR Mandville, PJM Kerstens. Algorithms for
confocal optical inspection system. Proceedings of the IEEE Workshop on Applications of Computer Vision, IEEE Computer Society Press, Los Alamitos, CA, 1992,
pp 298-305.
94. GJ Brakenhoff, HTM vander Voort, EA van Spronsen,N Nanninga. 3-Dimensional
imaging of biological structures by high resolution confocal scanning laser microscopy. Scanning Microscopy 2:33-39, 1988.
95. S Inoue. Foundations of confocal scanned imaging
in light microscopy. In: JB Pawley, ed. Handbook of Biological Confocal Microscopy. New York: Plenum, 1990,
pp 66-97.
96. Y Vodovotz, E Vittadini, J Coupland, DJ McClements, P Chinachoti. Bridging the
gap: use of confocal microscopy in food research. Food Technol 50:74-78, 1996.
Computer Vision
97.
98.
99.
100.
101.
102.
103.
97
BE Brooker. The study of food systems using confocal laser scanning microscopy.
Microscopic Anal 9: 13- 15, 1991.
AN Hassan, JF Frank, MA Farmer, KA Schmidt, SI Shalabi. Formation of yogurt
microstructure and three-dimensional visualization
as determined by confocal scanning laser microscopy. J Dairy Sci 78:2629-2636, 1995.
JCGBlonk,HvanAalst.Confocalscanninglight-microscopyinfoodresearch.
Food Res Int 26:297-311, 1993.
K Ding, S Gunasekaran. Three-dimensional image reconstruction for food microstructure evaluation using confocal laser scanning microscope. Vol.
12. Artificial
Intelligence Review. Kluwer Academic Publishers, 1998, pp 245-262.
S Gunasekaran, K Ding. Three-dimensional characteristics of fat globules in Cheddar cheese. J Dairy Sci 82(9):xxx-xxx, 1999.
CC Hoyt. Toward higher resolution, lower cost quality color and multispectral imaging. Advanced Imaging (April): 1-4, 1995.
BK Miller, MJ Delwiche. Peach defect detection with machine vision. Trans ASAE
34(6):2588-2597,1991.
104.
BPark, YR Chen, RWHuffman.Integrationofvisible/NIRspectroscopyand

multispectralimaging for poultrycarcassinspection.
J FoodEng 30 (112):
1996.
105. MJCohen,GHOlsen.Near-infraredcamerainspectswithclarity.LaserFocus
World. Pennwell Cop., Tulsa, Oklahoma (June):269-270, 1996.
106. JG Ibarra, Y Tao, J Walker, C Griffis. Internal temperatureof cooked chicken meat
through infrared imaging and time series analysis. Trans ASAE
1999.
107. MA Shahin, EW Tollner. Apple classification based on water core features using
fuzzy logic. ASAE Paper No. 97-3077, St. Joseph, MI, 1997.
108. S Dieterle. X-rays make sense. Resource (March): 13, 1999.
109. MW Eklund, G Ravichandran, MM Trivedi, SB Marapane, S.B. Real-time visual
1 IO.
111.
112.
113.
114.
115.
116.
117.
tracking using correlation techniques. Proceedings of the Second IEEE Workshop

on Applications of Computer Vision, IEEE Computer Society Press, Los Alamitos,
CA, 1994, pp 256-263.
K Liao, MR Paulsen, JF Reid. Real-time detection of colour and surface defects
of maize kernels using machine vision. J Agric Eng Res 59:263-271, 1994.
DW Lake. Proceedings of the111 Food Processing Automation Conference, ASAE,
St. Joseph, MI, 1994, pp 39-44.
B Ni, MR Paulsen, K Liao, JF Reid. An automated corn kernel inspection system
using machine vision. ASAE Paper No. 933032, St. Joseph, MI, 1993.
R Nelson.DSP-basedprocessingboardsboostoutput.VisionSystemsDesign.
Pennwell Corp., Tulsa, Oklahoma, December 1997, pp 40-44.
A Wilson.CPCI-basedframegrabber.VisionSystemsDesign.PennwellCorp.,
Tulsa, Oklahoma, March 1999, pp 73-80.
A Wilson. Image processing and neural networks classify complex objects. Vision
Systems Design. Pennwell Corp., Tulsa, Oklahoma, March 1999, pp 63-70.
KL Boyer, AC Kak. Color-encoded structured light for rapid active ranging. IEEE
Trans PAM1 9(1):14-28, 1987.
R Bachnak. 3D vision: An overview of techniques and application. Proceedings of
the Applied Machine Vision '94 Conference, SME, Minneapolis, MN, 1994.
98 Gunasekaran
and
Panigrahi
1 18. Choosing imaging boards and software. In: Product Hand Book. Data Translation,
Inc. MA. pp 236-245, 1996.

119. DA Krohn. Fiber Optic Sensors. Fundamentals and Applications. Instrument Society of America, NC. pp 1-10, 1988.
120.Designerschecklist.In:Schott-FostecStandardProductCatalog.Schott-Fostec,
LLC. Auburn, NY (June):33, 1999.
121. Benefits of the PC1 Bus for data acquisition and imaging. In: Product Hand Book.
Data Translation, Inc. MA. pp 246-247, 1996.
122. What color is color? In: Image Processing Catalog. Data Translation, Inc. MA. pp
8-20-8-23,1988.
123. Dictionary of terms. In: J Kaufman, ed. IES Lighting Hand Book. New York: Illuminating Engineering Society, pp 1-6, 1972.
124. JC Huber. getting down to the basics of fiber-optic transmission. R&D Magazine,
(February):l15-118,1994.
Delayed Light Emission and

Fluorescence
Suranjan Panigrahi
1.
INTRODUCTION
Biochemical degradation of chlorophyll results in exposure of carotenoids, giving

fruits and vegetables a characteristic maturity symptom of yellowing (1). Since
our eyes perceive colors, subjectivevisual evaluation has traditionally been used
to judge maturity and other quality attributes of fruits and vegetables. Recent
progress in the development of new methods utilizing optical properties has provided several objective techniques for rapid, accurate, more uniform quality assessment. The discovery of delayed light emission (DLE) by Strehler and Arnold
(2) has offered an alternative for evaluating quality
of chlorophyll-containing
plant materials.The strong dependenceof chlorophyll concentration and the duration and intensity of DLE have been applied in developing indices for different
(3). Recent investigations (4)
quality attributes of various fruits and vegetables
show that if the variables that are known to affect the duration and intensity
of
DLE are carefully controlled, the efficiencyof sorting fruits and vegetables could
be increased.
Fluorescence is another phenomenon exhibited by food and biological materials, Le., some materials fluoresce when irradiated with light. George Stokes
initially observed fluorescence in the early 1800s (5). Although this is one of the
as a tool in
oldest analytical methods, it has only recently gained importance
biological sciences related to food technology (6). Today, several quality evalua99
100
Gunasekaran and Panigrahi
tion applications employ fluorescence. Though fluorescence is similar to DLE in

terms of light emitted, the mechanisms are very different. By definition, DLE
refers to light emitted after excitation lasting up to several seconds, whereas
fluorescence ceases abruptly when the exciting energy source
is removed (6).
Also, unlike DLE, fluorescence can be artificially induced by adding some fluoto be sensitive,
rescence activators. Nonetheless, both techniques have proven
rapid, reliable, and reproducible.
In this chapter, the nature and occurrence
of DLE and fluorescence are
presented. The effects of various environmental factors are considered. Practical
of
applications of DLE and fluorescence measurements for quality evaluation
food and biomaterials are reviewed, and future research needs are outlined, along
with the instrumentation aspects associated with developing a measurement system for automatic quality evaluation.
II. THE PHENOMENONOFDLE

Strehler and Arnold ( 2 ) accidentally discovered the phenomenon of DLE while
studying adenosine triphosphate (ATP) formation during photosynthesis usinga
firefly luminescent system as an indicator. They observed that when green plants
are irradiated, they give off
light for a considerable period after illumination.
This phenomenon of DLE was explainedas a reflection of certain early reactions
in photosynthesis which, by virtue of their reversibility, are capable of releasing
a portion of their stored chemical energy through
a chemiluminescent mechanism-luminescence due to the energy liberated in a chemical reaction. Adding
nonpolar solvents (e.g., ethanol and ether)in moderate concentrations or irradiating chlorophyll solutions with ultraviolet (UV) light destroys the luminescence
a chemiluminescent mechanism
of intact cells, which supports the theory that
might be involved in DLE. However, the chemiluminescent mechanism involved
in DLEwasthoughttodifferfrom
themechanisminvolved
in phosphorescence-luminescence that persists after removal
of the exciting source due to
storage of energy. DLE was believed to involve biochemical reactions that produce excited molecules and to be associated with an enzyme system. Phosphorescence, on the other hand, is purely a physical reemission of trapped light energy.
Preliminary measurements of wavelength distribution of the emitted light
indicated a close correspondence to the wavelength distribution of chlorophyll
fluorescence. It wasshown,however, thattheluminescence of DLEismore
closely related to photosynthesis than to fluorescence (2,7). The similarities between the luminescent reaction and photosynthesis include: (a) the nature
of temperature dependence; (b) the rate at which the reactions are destroyed
by UV
light; (c) the range of saturation intensity; (d) chemical compounds that inhibit
101
the reactions; (e) suppression by carbon-dioxide; and (f) the drop in intensity
produced by continuous illumination with time.
Contrary to the interpretation of Strehler and Arnold (2) that the luminescence phenomenon is a consequence of the reversibility of some of the early
enzymatic photosynthetic reactions, the early processes following light absorption were observed to be nonenzymatic in nature (8,9). Later investigations suggested an interpretation of the physical processes leading to DLE and photosynthesis in terms of semiconductor theory (10,l I). The emitted light was attributed
to an electron transition between the first excited singlet state of chlorophyll and
the ground state (9,12). However, the luminescenceof DLE is an extremely complex temperature-dependent process, suggesting that a multiprocess mechanism
may be involved. Based on their temperature-dependent DLE studies on Chlorella, Scenedesmus, and spinach chloroplasts, Tollin et al. (9) showed that the
early processes following light absorption are purely physical, whereas the later
stages of emission are enzymatic in nature. Some investigators (13-15) have
suggested that DLE is regulated by the rate of the electron transport reaction,
and others ( 16- 18) have indicated that it is controlled by the rate of photophosphorylation (light-induced esterification of compounds with phosphoric acid).
Typically, induction of delayed light would show a very fast increase to
the initial level, then a slow increase to the peak, and a slow decrease toa steadystate level (2,7,19).
111.
FACTORSAFFECTINGDLE
All fruits, vegetables, and plant materials undergoing photosynthesis probably

produce DLE. Major factors affecting the intensityof emitted light are:(a) wavelength of excitation, (b) excitation intensity, (c) excitation time
and time after
illumination, (d) dark periods, (e) sample thickness
and area of excitation, ( f )
temperature, and (8) chlorophyll content of the plant material.
A.
Wavelength of Excitation
The DLE intensity produced from intact green peaches for

a broad activation
spectrum is shown on a relative scale in Fig. 1. Although the visible spectrum
peak is observed at about 680 nm, the fruit is significantly excited over a range
of about 625-725 nm. Significant DLE intensities were observed from oranges
when excited by white light (tungsten lamp) in the 680-740 nm range with a
peak at 710 nm. Experiments with tomatoes (20) andtea leaves andspinach
chloroplasts (21) exhibited similar results of peak DLE intensity at about
700
700
102
600
500
400
800
Wavelength of Excitation (nm)

Fig. 1 DLE activation spectrum for green peach measured 2 s after illumination. (Data
from Ref. 24.)
nm. However, the wavelength dependence may vary slightly based on the measurement system used.
B. ExcitationIntensity
In general, the intensityof DLE bears a direct relationship to the excitation intensity until saturation occurs. Itoh and Murata (7), however, observed an opposite
trend with spinach chloroplasts. Additional illumination has little effect in prolevel (22-24). For
ducing an increase in DLE intensity beyond the saturation
oranges, for example, DLE intensity levels off after
an excitation intensity of
2750 lx (Fig. 2 ) . This aspect is desirable for DLE measurements since small
changes in excitation will not interfere with quantitative measurements of DLE,
it
provided the illumination is sufficiently above the saturation point. However,
has been shown (7,24) that higher illumination intensity lowers the time required
for saturation. Total excitation energy required to obtain maximum DLE is differto obtain high DLE intensity
ent for each product. Optimal measuring criteria
for various fruits are listed in Table 1.
C. Excitation
Time and Time After Illumination
After a certain duration

of excitation, DLEmay be observed over varying periods.
At room temperature, DLE is observable for 0.25 ms or less up to one hour
(2,13). However, the intensity of DLE decreases with the time after excitation,
which is known as the decay of DLE. Various investigators have shown several
103
2000
4000
6000
Exciting Illurnination intensity(Ix)
Fig. 2 Effect of exciting illumination intensity on DLE from green oranges after different decay periods: A, 0.7 s; B, 1.0 s; C, 1.5 s; D, 2.0 s. (Data from Ref. 22.)
phases of DLE decay (9,13,14). The most rapid known has a half-life
of about
one ms and is popularly termed millisecond delayed light emission (19). Certain other systems have much longer half-decay times. The decay curves
of Chlorella luminescence are shown in Fig. 3. The decay appears to be faster at lower
temperatures than at higher temperatures. For a given decay period, maximum
DLE intensity is reached within the first 2-4s. Longer excitation times gradually
degrade DLE intensity. Therefore, it is advisable to provide short (about 2-4 s)
but strong excitation (to assure saturation). Temperature has a profound effect
Table 1 OptimalMeasuringCriteria
for ObtainingHighDLEIntensitya
Excitation
(min)
Temperature
Illuminationperiod
Product
Dark
Tomato
Satsuma orange
Persimmon
Japanese apricot
Banana
Papaya
Measured after 75-s delay.
Source: Ref. 38.
10
>20
15
(1x1
Time (s)
(C)
13-17
550
2750
2800
1-32
3-6
4-7
1-3
23-28
>20
>SO0
>IO
18-25
2750
>5500
15-22
1-2
2-4
104
1
'm
0.8
5 0.6
-E
W
&
0.4
0.2
0
0
Time After Excitation (9)
Fig. 3 DLE decay curves from Chlorella at 28C (A) and 6.5"C (B) (Data from Ref. 2 . )
on the shape of DLE decay. In general, the lower the temperature, the faster the
decay of DLE intensity. Jacob et al. (24), however, observed an opposite trend
of faster decayat higher temperature in immature oranges. The effect
of temperature on DLE is discussed further in a following section.
D. DarkPeriods
In order to eliminate the effect of previous illumination on DLE intensity, the
product is kept in a dark chamber fora certain length of time between successive
on the same product.
illuminations when more than one measurement is made
In some instances, the product may be subjected to dark periods after excitation
but prior to DLE measurement. Based on the studies of DLE in tomatoes (20),
Satsuma oranges (22), and tea leaves (21), definite relationships between the
preconditioning dark periods and DLE have been established. Usually, short dark
periods (<1 min) cause a reduction in DLE intensity while longer dark periods
(about 10-15 min) bring about some recovery of DLE intensity. Regarding chlorophyllderived from spinach chloroplasts, the illumination after varying dark
periods showed initial DLE peaks of varying levels depending on the duration
of dark periods (7). The effect of a dark period before excitation on DLE intensity
of yellow persimmons excited at 5500 Ix is shown in Fig. 4.
E. Sample Thickness and Area of Excitation

Scattering and absorption of light limit the depth below the surface, which contributes to DLE. This may be considered a disadvantage in DLE applications to
detect subsurface defects. Jacob et al. (24) found 2.5 mm as the limiting depth
105
10
20
Dark Period (rnin)
Fig. 4 Effect of dark period before excitationon DLE intensity from yellow persimmons
excited at 5500 lx as a function of decay periods: A, 0.9 s; B, 1 .0 s; C, 1 .5 s. (Data from
Ref. 24.)
for DLE detectionin tomatoes. Chuma and Nakaji (20) obtained highDLE intenrind, the interior being scooped
sity from tomato flesh 6 mm thick below the
out.
A direct relationship between the area of excitation and the intensity
of
emitted light has been observed (20,22,24). Figure
5 represents the linear relationship between the area of excitation and the DLE intensity of green tomatoes at
different decay periods. Also, a larger area of excitation resulted in a faster DLE
decay. Therefore,it is necessary to carefully define the areaof excitation for each
product and for each quality attribute to obtain reproducible results. This would
have a bearingon developing automatic grading systems for fruits and vegetables.
F. Temperature
TemperatureeffectsonDLE,asobservedearlier,areextremelycomplex
to
(9,22,25). Chlorella luminescence shows a temperature dependence similar
that of an enzyme-catalyzed reaction (Fig. 3). The amplitude and decay kinetics
of DLE are temperature-dependent in the time range of seconds (3,8,9,15). DLE
2
3
4
5
Illumination Area (cm)
Fig. 5 DLE intensity vs. area of excitation in green tomatoes excited at 5500 Ix. (Data
from Ref. 20.)
intensity is generally maximum in a narrow range of sample temperatures. However, the variation from maximum DLE intensity is dependent on the particular
of DLE from green
product. Figure 6 represents the temperature dependence
tomatoes and bananas, respectively. Peak intensities are at 16 and 25C for green
tomatoes and green bananas, respectively.Also, higher temperatures havea more
profound effect on DLE from tomatoes than from bananas. The temperature dependence of DLE, however, is not a major problem when agricultural products
encounter relatively small temperature changes (24). Abbott and Massie (26) related temperature and duration of exposure to chilling temperaturesof cucumber
fruit to DLE; Chuma et al. (22) observed an increase in DLE from orange peel
up to 30C and a sharp decline reaching almost zero level beyond 40C; and
Nakaji et al. (21) found a strong dependence of DLE on the temperature of tea
leaves. Terzaghiet al. (27) reported that the temperatureof maximum DLE upon
chilling was strongly correlated withlateral phase separation temperature but was
about4Cloweronaverage.Thephenomenonobserved
in chilling-sensitive
plants during chilling a minor component below freezingis lateral phase separation. These investigations indicate that large temperature changes have a definite
effect on DLE, which can be used to predict certain quality attributes.
107
.. .
3.5
n
10
15
25
20
30
35
Temperature ("C)
Fig. 6 Effect of temperature on DLE intensity for green tomatoes

(A) and green bananas
(B) after a decay period of 0.7 s. (Data from Refs. 20, 23.)
G. ChlorophyllContent
DLE is probably produced by all vegetables, fruits, and plant materials undergoto the chlorophyll concentration of the
ing photosynthesis and can be related
product (24). This is perhaps the most important aspect of DLE from a practical
viewpoint because the change in the level of chlorophyll content is a primary
20
40
60
80
100
Time (s)
Fig. 7 DLE intensity at different stages of maturity of lemons: A, Dark green: B, light
green: C, silver, D, yellow. (Data from Ref. 24.)
Table 2 Empirical Relationships Developed for DLE from Various Products as a Measure of Chlorophyll Content or Peel Color
Based on Chlorophyll Content
0
Q)
Correlation
Product (variety)
Empirical relationship
+ 0.298
(W
Measurement conditions
Ref.
content (pg/
0.92
Illumination intensity = 5500 lx

Illumination area = 9.07 cm?
Excitation time = 4 s
Decay period = 0.7 s
Dark period = 10 min
Product temperature = 13OC
Source = 100 W tungsten lamp
Product temperature = 20C
Source = Xenon flash Illumination
area = 4.425 cm2
Excitation time = 0.4 ms
Product temperature = 27C
Source = 80 W tungsten halogen lamp
Illumination intensity = 5500 Ix
Product temperature = 18.5OC
Illumination intensity = 5500 Ix
Illumination area = 4.4 cm'
20
Dependent variable
C-chlorophyll
I00 cm')
Tomato (Japanese)
DLE = 0. I63 C
Tomato (Duke)
DLE = 0.0121 C
0." 13
C-total chlorophyll content

(pg/g fresh weight)
0.86
Tea leaves (Japanese)
DLE = 0.0896 C +
0.318
C-chlorophyll content (pg/

100 cm? leaf area on
hack side of leaf)
0.86
Papaya (Hawaiian)
DLE
P-papaya maturity grade

(P = 9 for mature green;
P = 28 for 3/4 ripe)
0.92
Orange (Satsuma)
DLE = (0.235 C 0.276)"' + 0.512
Banana (monkey)
DLE
DLE intensity, V .
0018 P
1.054
DLE = 10.9 exp(-0.72

GPC)
C-chlorophyll content of
peel (pg/100 g fresh
weight)
GPC-grade
of peel color
21
29
22
P
v)
FP
23
a
a
a
'CI
1.
(CI
109
indicator of maturity and sometimes quality

of various fruits and vegetables.
Jacob et al. (24) observedthat DLE related to apparent chlorophyll concentration
in apples,apricots,bananas,nectarines,
bell peppers,olives,onions,oranges,
peaches, persimmons, plums, pomegranates, and tomatoes. Figure 7 depicts the
relationship between DLE and four commercialcolor grades of lemon from dark
green to yellow, representing decreasing levels of chlorophyll concentration in
the lemon skin. Table 2 lists empirical equations relating DLE intensity and the
chlorophyll content or peel color based on chlorophyll content obtained for various products. Since the measurement conditions have a pronounced effect on
DLE intensity, these empirical relationships must be carefully interpreted and
used.
IV.
QUALITYEVALUATIONBASEDON DLE
Typically, the process of ripening involves degradation of chlorophyll, which

exposes carotenoids, giving a characteristic maturity symptom of yellowing ( I ) .
Watada et al. (28) observed that chlorophyll content decreased steadilyfrom 13.4
0 to 46.7 pg/g
to 0.3 pg/g fresh weight, and lycopene content increased from
as tomatoes ripened from stage 1 to stage IO. Similar changes in pigment content
of tomatoes and papayas were also observed as these fruits matured (2.29).
Because of the change in pigment composition, fruits exhibit reduced absorbance i n the 670 nm region as they mature. Spectrophotometric methods take
advantage of this optical characteristic in indicating maturity and other quality
attributes of fruits and vegetables. Because of their simplicity, spectrophotometric
methods have been widely used for maturity and quality evaluationof biological
materials.However,DLEmeasurementshavethefollowingadvantagesover
spectrophotometric methods (24):
1.
Illumination need not be spectrally well defined, which permits wider

choice and simpler design of the exciting source.
2. Illumination and measurement are not simultaneous, i.e., illumination
can be removed in time and position from measurement.
3. Beyond saturation level, additional illumination has little effect on increasing DLE intensity,so small changes in illumination will not interfere with quantitative measurement of DLE.
On the other hand. DLE measurements provide no compensation for size.
A large specimen emits more light than a smaller one of equivalent chlorophyll
concentration. To reduce this effect, usually only a portion of each specimen is
examined. The disadvantage of this is getting a result that does not characterize
the maturity of the whole product. This can be overcome, however,
by special
110
measurement techniques such as taking the average

of several readings on the
same specimen at different places.
Recent advances in automatic product quality evaluation have been made
in the area of computer vision systems. Computer vision
or image processing
systems essentially duplicate conditions as the eye sees an object but are able to
enhance or statistically evaluate some aspect
of an image thatis not readily apparent in its original form. Image processing applications to date include detecting
bruises in apples (30), stem and blossom ends in tomatoes (3 l), orientation and
shape of bell peppers (32), and various quality factorsof corn, rice, and soybeans
(33-37). However, almost all of these applications rely largely on the contrast
between defective and undamaged regions
of the image. Therefore, an image
processing system may not be able to detect subtle color changes associated with
various stages of ripening and maturity.
Jacob et al. (24) were among thefirst to investigate the application of DLE
to quality evaluation of fruits and vegetables. They observed significant differof lemon from dark green to ripe,
ences among four commercial color grades
representingvariouslevels
of chlorophyllconcentration
in thelemonskin
(Fig. 7).
Chuma and Nakaji (20) observed an almost linear relationship between the
chlorophyll content of tomatoes and DLE intensity. Higher efficiency was
reported in sorting green tomatoes than red ones. Since separating green tomatoes
from breaker stage fruits is of greater importance, such measurement could be
of commercial significance. Forbus et al. (4) classified tomatoes into three stages
of maturity on the basis of visual color and evaluated for differences in DLE
intensityandchlorophyllandcarotenoidpigmentcontents.Resultsshowthat
DLE has a high potential for use in the automated sorting of fresh market tomatoes into categories of maturity. Jacob et al. (24) reported sorting tomatoes under
field conditions based on DLE measurements.
DLE characteristics as related to the quality of Satsuma oranges were also
investigated (22). Sorting efficiencies of 55, 90, 84, 86, and 100% were reported
based on DLE measurements for five color groups of dark green, light green,
color groups
yellowish green, yellow, and orange. Since identifying the latter
are of more practical importance, the corresponding higher efficiencies indicate
applicability on a commercial scale. The usefulness of DLE measurements for
commercial color grading is further enhanced by the fact that surface treatments
on oranges such as brushing and waxing have negligible effects on DLE (22).
However, storage treatments seem
to have a certain effect on DLE. Based on
investigations at room temperature, cold storage (2"C), and dark storage, Chuma
et al. (22) found a decline of DLE up to 336 hours, with the fastest decrease for
room temperature-stored product.
Measurements of DLE on bananas indicated that the intensity of emitted
light decreased with ripening, and bananas just beyond optimum ripening had
111
only a trace of DLE (23). The grade of peel color of bananas had been related
to DLE intensity (Fig. 8), which indicates the possibility
of using DLE measurements to determine the state of maturity of bananas. Other quality attributes of
bananas such as sugar content and firmness were also related to DLE. As the
sugar content of the flesh (S, Brix) of bananas increased from 1 to 20% during
ripening, the intensity of emitted light (Y, volt) decreased almost linearly as:
-0.773
(1) S
+ 3.26
Similarly, the relationship between the fruit firmness

was expressed as:
Y(2)= 14 F
(F, N) and DLE intensity
3.50
DLE intensityalso has a significant correlation with firmness for other fruits
such as apricots and papayas (38).
DLE characteristics of the chlorophyllof green tea leaves werealso investigated to serve as a means of nondestructive estimation of the quality of tea and
tobacco leaves (38). Each ingredient valueof leaf color (hue, value, and chrome)
wasfound to correlate highlywith DLE intensity of fresh leaves. The linear
relationship between the chlorophyll content of fresh leaves and DLE intensity
could be used to estimate the quality of tea leaves.
Abbott and Massie (26) and Abbott et al. (39) related the temperature and
duration of exposure to chilling temperatures of cucumber and bell pepper fruit
and coleus to DLE measured at nonchilling temperatures. They observed that the
DLE measurement would help determine the mechanisms of chilling injury and
the methods of ameliorating chilling injuryof cucumbers, bell peppers, and other
chlorophyll-containing tissues (40). Besides chilling injury, DLE measurements
can also be used to inspect fruits and vegetables for mechanical stress during
Color Grade
Fig. 8 Banana peel color grade vs. DLEintensity.(Data
from Ref. 23.)
112
harvesting (41) and injury from sunburn or air pollution before visible symptoms
develop (42). Based on the effects
of quarantine heat treatments on DLE measurements on papayas (43), Forbus and Chan (44) investigated the potential for DLE
as a biochemical indicator of papaya heat treatment. The DLE system was not
as sensitive to deleterious effects
of heat on the ethylene-forming enzyme system.
However, they concluded that DLE measurements provided a quantifiable, nondestructive means for measuring the primary deleterious effects of heat damage.
Forbus and Senter(45) evaluated Saticoy cantaloupes at six stagesof maturity for differences in DLE, chlorophyll, and soluble solids. DLE varied significantly and predictably by stage of maturity, indicating its potential as a rapid,
nondestructive method of measuring cantaloupe quality.
Forbus et al.(29) developed an index for separating ripe and unripe papayas
(Fig. 9). Working with papaya maturity ranks, they found that ranks 9 through
28 (mature green through three-quarters ripe4 days after harvest) couldbe sorted
more accurately than ranks 1 through 9 (immature green 4 days after harvest).
Half-ripe or riper papayas (maturity ranks21 through 28) are the most susceptible
in
to fruit fly infestation(43).Therefore,DLEmeasurementscangreatlyaid
identifying and removing infestation-prone fruits and thus assure quality in large
for DLE
commercial shipments. Forbus and Chan (44) demonstrated the potential
0.9
I
.b
u)
c
0.8
Q)
c
m
0.7 Overhalf-ripe
K
UI
.-
Under half-ripe
w 0.6
21
n
25
0.5
27
0.4
5
10
15
20
25
30
Relative Maturity Rank
Fig. 9 Regression of DLE intensity vs. papaya maturity (I' = 0.92). Nutnhers 9 through
27 on the tigure represent papaya maturity rank arranged in order of increasing tnaturity.
(Data from Ref. 24.)
113
as a rapid screening technique for detecting papayas ripe enough

to be susceptible
to fruit fly infestation. There was a high correlation ( r = -0.92) between DLE
intensity and Hunter b values for freshly harvested papayas at seven stages of
maturity. Recent reports (46-48) indicate a significant decrease in DLE intensity
with increased maturity grade of different peaches. Similar results were observed
for Japanese persimmons (49), plums(50),and muskmelons (51). DLE also corof canary melons (cv. Juan
related highly ( r = 0.85) withthematurityindex
Canary) (52). The maturity index was calculated based upon values for chlorophyll, yellow pigments, and soluble solids content.
V.
INSTRUMENTATION AND MEASUREMENTUNITS
There have been varying designs of DLE measurement units to evaluate fruit
and vegetable quality (20,24,29,43). Jacob et al. (24) reported one of the earliest
instrumentation systems for continuous DLE measurement and sorting fruits. Forbus et al. (53) developed an experimental DLE meter for horticultural crops.
According to them, DLE measuring systems must incorporate the following components: (a) a light source for illuminating a sample
to activate DLE, (b) a totally
dark enclosure in which to isolate the sample for measuring DLE after illumination is removed, (c) a photomultiplier tube mounted inside the totally dark enclosure for detecting light emitted from the sample, (d) an amplifier for applying
of emitted
low-intensity signals, and (e) a recorder for recording the intensity
light as a function of time.
The basic components of instrumentation and measurement units for autoto most optical and/or
matic quality evaluation based on DLE are very similar
spectrophotometric units. A detailed description of various detector properties
and specifics of associated electronics to measure photoluminescence and other
optical radiations can be found in an excellent multivolume treatise on optical
radiation measurements (54-56). However, an attempt is made here to broadly
unit. The overall
outline various components of a typical optical measurement
measurement unit can be divided into four major systems
to facilitate: sample
illumination, viewing and sensing, signal processing, and sorting.
A.
IlluminationSystem
Incandescent lamps (tungsten filament and tungsten halogen), arc lamps, andfluorescent lamps have commonly been used as sources
of radiation in optical instruments. Occasionally, low-power lasers have also been used as
light sources in
optical systems (57,58). Laser light is highly intense, coherent and directional,
has a well-defined beam diameter, and
is readily adaptable for optical use (57).
However, its use may be limitedby its monochromatic nature. Primarily, a source
114
Panigrahi
and
Gunasekaran
should emit a substantial amount

of radiation in the spectral regionof interest. For
more general applications, light sources must have a wide band
of illumination of
sufficient intensity. However, because of the broad activation spectrum of DLE,
the wavelength of illumination is not very critical for most measurement conditions where the products surface layer is the primary indicator
of quality because
DLE reaches saturation at some threshold illumination levels. DLE intensity
is
usually large in the range of 650-700 nm wavelengths. Readily available illumination sources such as fluorescent and incandescent lamps produce good output
within this range. However, a very large increase
in illumination is needed to
penetrate enough to reach deep tissues containing chlorophyll (23).
Glass color filters that absorb in certain wavelength regions and transmit
in others are commonly used in optical instruments for wavelength control (5860). Using a prism or grating to disperse a light beam into a spectrum and using
a narrow slit tolet only the desired wavelength band pass through can also isolate
light within a certain wavelength band. Beam splitters and optic fibers also serve
to control the radiation. Use of beam splitters in optical systems adds the convenience of having two light beams from a single source (57), which can help in
making dual measurements and allow for a more compact system. Optic fibers
provide easy access for illuminating and sensing of light over hard-to-reach and
to a bare minimum.
intricate areas. Also, light transmission losses are kept
B. SampleViewingandSensingSystem
The sequence of operations in this system
is singulation, conveying and orienting,
and scanning and sensing. The singulation stageis incorporated into systems that
of discrete objects is formed
evaluate individual products. At this stage, a bulk
in a noncontiguous single file. Even though there is no established theoretical
basis for the study of the process of singulation, Henderson and Shawver (61)
described singulation as a three-stage process. First, the objects received in bulk
are spread into a disordered monolayer. Second, the monolayer is organized into
rows. Third, the objects in the rows are separated to a specified spacing between
them.Acompendium
ofsingulationandassociateddevicespublishedby
Shawver and Henderson (62) is an exhaustive reference on this subject. Conveying the singulated materials at the proper orientationis critical for an accurate
evaluation of quality. Since most agricultural and biological products are irregularly shaped, orienting them properly is not easily achieved. An orienting and
conveying device for sorting dried prunes developed by Burkhardt and Mrozek
(63) is a good example. Gaffney (64) suggested incorporating techniques
that
measure the size and shape of objects into the design of a conveying and orienting
unit. Another important aspect is the speed of conveying, which could affect the
sorting efficiency.
115
Selecting a suitable sensor unit and associated electronics should be done

judiciously with due regardto the system as a whole. The choice of photodetector
should be based on certain important parameters: physical dimensions, sensitivity, speed of operation, and noise characteristics. Some popular optical detectors
A photomultiplier
arephotomultipliertubes,photoresistors,andphotodiodes.
tube is a fast, efficient detector and is widely used in the near-UV, visible, and
near-infraredregions.However,theyarerelativelylargeandrequireahighto their cost and somewhat
difficult
voltage power supply, which contributes
mechanical design (64). For these reasons, useof solid-state sensors such as photoresistors and photodiodes has become common. A photoresistor is essentially
a film of material whose resistanceis an inverse function of incident illumination.
Photoresistors are inherently slow. Because their sensitivity is somewhat dependent upon past history, they are not entirely consistent. When photodiodes are
used in reverse-biased mode, they respond very quickly and linearly over many
decades of irradiance.
a wide
Generally,opticalsensorsshouldhaveaspectralresponseover
wavelength band. They should also exhibit a flat response at the wavelength of
interest rather than a peak response for stable output. The peak DLE emission
a good
being in the red region (650-700 nm) indicates that a phototube with
response in the red region
is the most suitable. Besides matching the spectral
aspects of incident radiation, the selection of photodetectors should also consider
geometric aspects (e.g., angle of incidence, beam cross section, density and uniformity of radiation, relative position of the sample) and temporal aspects (e.g.,
rise time of pulsed radiation as compared to rise time of the detector)of the input.
An important difference between common optical systems and DLE measuring units is the provision of dark chambers in the latter. These dark chambers
may be placed to provide a dark period between illumination and measurement.
In the case of multiple measurements on the same sample, dark chambers can
provide dark periods between subsequent illuminations. The length of the chamber and/or the speed of conveying the product should be carefully controlled to
allow for a sufficient dark period resulting in optimal DLE.
C.SignalProcessing
System
Signal processing is an essential part of the measurement unit. Often the ability
of the signal processing componentwill dictate the designof the rest of the system
(39). In a broad sense, signal processing simply means transforming the photodetector response into a more useful form to aid in decision making. For this, it can
or
be as simple as direct reading of the electrical signal coming from the detector,
it can be as involvedas requiring sophisticated mathematical analyses ofmeasurement data performed with the aid
of large digital data processing facilities.
In
116
Panigrahi
and
Gunasekaran
processing photoluminance or fluorescence data, simple manipulation of measured values from some meter or analog-to-digital converter
is inadequate. In
this case, the domainof data analysis may extend from the direct detector output
to the digitization process and
finally to the conversion of these quantities to
information of analytical interest (65). For this reason, greater reliance on signal
processing is a necessity. Typically, a microcomputer will be dedicated for this
with specially developed software and hardware incorporated into the system. In
general, this system provides overall control, calibration, and operation
of the
unit and helpsto display, record, and/or store
the optical characteristics measured.
D. Sorting System
At this stage, this is the actual classification according to the information gained
from the signal processing system. This may be grouping fruits or vegetables
into different grades or rejecting defective samples. Prompted by the microcomputer signal,a solenoid could actuate a deflecting vane (24)or an air gun (62,66).
of
A pneumatically operated sliding gate could also be used to direct the flow
defective products (66,67). Sorting speed should match the conveying speed, and
to cause mechanical damage to the
theunitshouldbecarefullydesignednot
objects being sorted.
VI.
FLUORESCENCE
According to Guibalt, Fluorescence refers to the property

of light (luminescence)emittedbymolecules
of a particlewhentheyareexcitedbyphotons
(light) (6). The word fluorescence is derived from the mineral fluorite
( 5 ) . Fluorite generally glows in blue color under sunlight and the UV spectrum
in the sunlight causes this. This process was observed by Sir George Stoke
in
early 1800s ( 5 ) and this phenomenon was named fluorescence (5). However,
fluorescence phenomenon is not just observed under UV light, it can also be
NIR spectrum of electromagnetic radiation
observed under visible, x-ray, and
( 5 ) . The emission (luminescence) can also be observed not only in visible, but
in NIR and infrared region( 5 ) . Thus, a general working definitionof fluorescence
canbewritten as theproperty of anobject(particle) to emitphotonswith a
wavelength (emission wavelength) higher than that at which the object was excited with (excitation wavelength) (5,6,104).It is to be noted that the fluorescence
phenomenon stops abruptlyas soon as the excitation sourceis removed (5,6,104).
For some materials, the emission (luminescence) continues after the removal of
excitation source. This process is called phosphorescence (5).
Guibalt (6) describes fluorescence as follows. Every molecule possesses a
series of closely spaced energy levels and can go from a lower to higher energy
117
Excited Trlplet State

Energy by
First Exc
SI -
v=4
v=3
v=2
so -
lnteratomlc Dstance Along

Critical Coordinate
Ground Electronic State
Fig. 10 Schematic energy level diagram for a diatomic molecule. The potential energy
level of a diatomic molecule are indicated with the various vibration levels represented
as v = 0. I , 2, 3, and 4 on each curve. (Reprinted from Ref. (6). p. 3.)
level by the absorption of a discrete quantum

of energy, e.g., light, equal in energy
to the difference between the two energy states (Fig. IO). Only a few molecules
are raised to a higher excited state and hence are capable of exhibiting luminescence. Between each main electronic state are the various vibration levels of the
molecules. When a quantum of light impinges on a molecule, it is absorbed in
about
s, and a transition to a higher electronic state takes place. This absorption of radiation is highly specific, and radiationof a particular energy is absorbed
only by a characteristic structure.The electron is raised to an upper excited singlet
state, S I , S I , etc. These ground-to-singlet transitions are responsible for the visible- and UV-absorption spectra observed for molecules.
The absorption transitions usually originate in the lowest vibrational level of the ground electronic
(1 0-j s), some
state. During the time the molecule spends in the excited state
energy in excess of the lowest vibrational energy level is rapidly dissipated. The
lowest vibrational level (v = 0) of the excited singlet state S is attained. If all
the excess energy is not further dissipated by collisions with other molecules,
the electron returns to the ground electronic state with the emission
of energy.
This energy emission is called fluorescence. Because some energy is lost in the
a
briefperiodbeforeemissionoccurs,theemittedenergy(fluorescence)has
longer wavelength than the energy that was absorbed.
118
When excited with a specific form of electromagnetic radiation (such as

UV light, NIR light), a part of the radiatiodenergy is reflected, some portion is
absorbed and the other portion
is transmitted. The exact amount of reflection,
to object. For some materials,
absorption and transmission varies from object
the absorbed radiationlenergy affects the atomic structure of the object (6).The
absorbed energy is generally expressed in integral units called quanta (6).The
quanta-energy relationship is expressed as:
E = -hc
h
where E = energy (J), h = Plancksconstant (6.63X
J . s), c = velocity
of light = 3 X 10 (m/s), and h = wavelength (m) (6).
The energy of the photons in the visible spectrum is 146-607 KJ/mol and
they cause electronic transition (6).The resulted fluorescence lasts for typically
I O ns (6). The time for which the fluorescence phenomenon exists is called fluorescence lifetime (6,104).
Fluorescence has two characteristic spectra-an excitation spectrum and
an emission spectrum. The excitation spectrum refers to the relative efficiency
of different wavelengths of exciting radiation to cause fluorescence. The shape
of the excitation spectrum should be identical to the absorption spectrum and is
independent of wavelengthof fluorescence measurement. The emission spectrum
represents the fluorescence spectral data.
It is the relative intensity of emitted
radiation at various wavelengths (6).
A.
Types of Fluorescence
Fluorescence may be classified as either primary or secondary. Primary fluoresto the intrinsic or native fluorescence, also known as autofluorescence, refers
cence characteristics of an object/product that fluoresces under UV or other electromagnetic radiation. Many plant products, botanical tissue components, animal
tissues including meat and fish products, and other food products possess autoSecondary fluorescence, also known as influorescence characteristics (6,l 13).
duced fluorescence, refers to the fluorescence characteristic of an object obtained
byaddingspecificfluorescenceactivatorscalledfluorochromoorfluorescent
stains or fluorophores. Some nonfluorescent objects show fluorescence characteristics after undergoing a chemical reaction or after the addition of chemical reagents (104,113).
In some cases, fluorescence may be classified as (6):
1.
Stokes fluorescence-thereemission
of lessenergeticphotons
that
is normally
have a longer wavelength than absorbed photons, which
observed in solutions.
119
2. Anti-Stokes fluorescence-emission at shorter wavelengths than those

of absorption, which occurs when thermal energy
is added to an excited
state or when a compound has many highly populated vibrational energy levels. It is often observed in dilute gases at high temperatures.
3. Resonancefluorescence-thereemission
of photonspossessingthe
same energy as the absorbed photons. It occurs in gases and crystals
and is not observed in solutions.
The relationship of fluorescence to the molar concentration of a material
is described by:
ql,, ( I
e)-"k
(4)
where q = quantum efficiency, I,, = radiant incident power, a = molar absorptivity, b = path length of the sample, and c = molar concentration (6).
The quantum efficiency or yieldq is furtherdefinedasthe
ratio of the
emitted to the absorbed number of quanta (6). The higher the values of q, the
I,, could cause
greater the fluorescenceof a compound. However, higher intensity
photodecomposition of the sample(fluorophores),thusreducingfluorescence.
Therefore, a source of moderate intensity is used (e.g., mercury or xenon lamp)
(6). For very dilute solutions, Eq. (4) reduces to Beer's law:
KqIoabc
(5)
where K is a constant (6).

The plot of fluorescence versus concentration should be linear at low con(6). Generally, a linear
centrations and reach a maximum at higher concentrations
response could be obtained until the concentration of the fluorescent species is
large enough to absorb a significant amount of incident radiation. To obtain a
linear response, the sample should absorb less than 5% of the incident radiation
(6).
Some food and biological products can be heterogeneous, and could contain
bigger particles (104). These could scatter light. This light scattering
is called
Raleigh's scattering (104). The absorbance from this scattering is inversely proportional to the fourth power
of excitation wavelength, h (A a l/h4), and can
be identitied as background absorption (104).
B. AdvantagesandLimitations
Fluorescence and phosphorescence are well suited as analytical tools for food
quality evaluation becauseof their extreme sensitivity and specificity. Concentration of substances as low as I in 10"' can be detected. The main reason for this
extreme sensitivity is that the emitted light is measured directly and can be increased or decreased by changing the intensity of the exciting radiation. The
120
specificity of fluorescence is due to the fact that there are few components that
fluoresce compared to those that absorb light energy, and two wavelengths are
used in fluorometry compared to only one in spectrophotometry. Many biological
and food products tend to have unique fluorescent characteristics. Two compounds that absorb radiation at the same wavelength will probably not emit at
the same wavelength. The difference between excitation and emission peak could
range from I O to 280 nm (6).
There are, however, the following limitations
of fluorescence as an analytical tool (6):
1.
It is strongly dependent on environmental and physicochemical factors,

such as temperature, pH, ionic strength. and viscosity.
2. Photochemicaldecompositionordestructionofthefluorescentcompound providing a gradual decrease in fluorescence could also create
a problem in fluorescence measurement. Methods used to avoid photochemical decomposition include using longest-wavelength radiation to
to strike the
excite the sample, not allowing the excitation radiation
sample for a long time, and avoiding or minimizing exposure
ofa
photochemically unstable standard solution to sunlight or UV prior to
the experiment/measurement process.
3. Quenching: the reduction of fluorescence by a competing deactivating
process is known as quenching. It may be observed due to oxygen,
impurities, temperature, and concentration. Dissolved oxygen, small
amounts of iodide, and nitrogen are very effective quenchers. Impurities such as dichromate also interfere with fluorescence. Temperature
has been found to reduce fluorescence by 1% (and up to 5 % ) per 1C
increase. High concentration of the fluorophore studied is known to
decrease fluorescence intensity.
C.Detection
of Fluorescence
Fluorescence can be measured by either transmission or reflection techniques.

The transmission technique is more subject to quenching effects than the reflection technique (6,104). Therefore, reflection measurements are more common.
Reflected light can be detected by different
instrunlents/systems such as spectrofluorometer, spectrophotometer, and imaging systems specially configured for
fluorescence measurements. The basic components of any fluorescence measurement system are (6,104):
1.
A light source to generate UV radiation in the required excitation wavelength.

2. An excitation filter toallowonlytheexcitationwavelengthspecific
121
radiation to interact with the sample. (In the cases where laser is used
as the light source, a filter may not be necessary.)
3. A detector and other associated electronics including A/D converter.
The detector will detect the emitted spectrum/fluorescence. The A/D
converter and associated electronics will convert the emitted signal into
equivalent digital form.
4. A host computer, which could host the detector assembly and store the
digitizedfluorescenceinformationforstorageandadditionalprocessing.
5 . Absorbance filter to allowtheemittedsignal
in theselectivewavelength range before the signal is digitized.
1. Illumination Sources
Manybroadbandlightsources
(e.g., xenon lamps, UVlight,mercurylamps,
xenon-mercury lamps, and lasers) are used as illumination sources (6,104). High
intensity discharge lamps such as mercury and xenon lamps are mostly used as
illumination sources (6,104,105). Mercury lamps provide sharp spectral output
and many times are used for calibration purposes along their expected lifetime
in high- and low-pressure format. Xenon
in average (105). They are also available
lamps provide continuous spectral output, generally from 270-400
nm (UV zone)
(104). Xenon gas is contained under high pressure in xenon lamps (104,105) and
therefore should be handled carefully. Xenon and mercury combination-based
(104). It is
xenon-mercury arc lamps provide higher output than a xenon lamp
always recommended to obtain detailed spectral output from the manufacturer
before selecting a particular light source.
Promising research is being carried out by various entities, and it is anticipated that UV light emitting diodes (LEDs) will be commercially available soon.
of the unique propertiesof
Another typeof illumination source is laser based. One
lasers is their ability to emit high-energy coherent lightof a specific wavelength.
of laser
Depending on the required wavelength and cost, the appropriate type
can be selected.
2. Detectors
Detection of fluorescent spectra can be performed using eithera photo-multiplier

tube (PMT) or photodiodes/phototransistors.
Though recent technological advancements have allowed the use of differin many fluorometers
ent types of detectors, PMT-type detectors can be found
(6,104). The PMT provides an output signal (in terms of current) that is proportional to the light received by PMT. However, it only provides an average signal
without any spatial information in it (6,104).
The exact type of photodiodes/phototransistorsselected for detecting fluo-
122
rescent spectra will depend on the type of detecting material used in the detector
(which is dependent on emission wavelength, active area of the detector, noise
equivalent power, and detectivity) (106). For fluorescence (emission) measureGe (germaments in the visible and near-infrared range, generally Si (silicon),
nium), and InGaAs(indium-gallium-aresenide) type detectors are used (107). For
of light is low, new types of
many fluorescence measurements where the level
photo-receivers such as femtowatt receivers (New Focus, Inc., CA) can be very
of food products, nanosecond phouseful. For fluorescence lifetime measurement
todetectors (New Focus, Inc., CA) with typical rise time of 1 ns may be appropriate (108).
To obtain fluorescent images, different types
of cameras/computer imaging
systems are used. A computer-based fluorescence imaging system differs from
a standard computer imaging system mostly by the image acquisition system or
camera. The camera oftenis incorporated with a microscope. The most commonly
used camera for a fluorescence imaging system is an intensified charge coupled
device (ICCD) video camera. The detectors on these cameras are cooled to provide high sensitivity and spatial resolution with video rate (30
frameds). Other
cameras that are used for fluorescence imaging use cooled and UV-enhanced or
back-illuminated CCD detectors (109). UV-enhanced or back-illuminated CCD
provides enhanced sensitivity in the UV or visible band to acquire low-light fluorescence images ( I 09). Fluorescence images can also be obtained with digital
signal processing (DSP) integrated CCD camera system. DSP circuitry allows
the integration of information over a user-defined amountof time (generally from
0.01 s to several minutes) to obtain high-quality images. This type
of system
works well with applications where the amount of light (fluorescence emission)
a
is relatively high. Fluorescence imaging can provide more information than
spectrofluorometer such as spatial distribution of intended fluorescence information. Thus, fluorescence imaging can be very useful for food quality measurements.
VII.APPLICATIONS
FOR FOOD QUALITYEVALUATION
Many food, agricultural, and biological materials possess autofluorescent characto fluoresce (i.e., secondary fluorescence) by
teristics, and many can be made
in a number
adding appropriate fluorophores. These properties have been used
of quality evaluation applications.
A.
AutofluorescenceTechniques
Separation of bran (pericap, aleurone, and fatty germ) from wheat flour (endosperm) is performed in many wheat-processing operations. Chemical analysis of
123
ash along with the color check are conducted to measure the effectiveness
of
separation (1 IO). However,thismethod is verytediousandtime-consuming.
Munck (1 IO) proposed a rapid methodin which the autofluorescence characteristics of different components of wheat seeds were used. He showed that the aleu350 nm wavelength light,
rone autofluoresces at 420 nm (blue) when excited with
and the pericarp fluoresces at 540 nm (yellow) when excited at 450 nm ( 1 IO).
A similar method was usedto study the components of sorghum via fluorescence
microscopy (68).
Mycotoxins (toxic fungal metabolites) in food products have been linked
with different forms of fatal and chronic illnesses for humans and animals (70).
Theyhavebeenreportedtobefound
in manyfoodproducts
at any stage of
production (70). Wood and Mann (70) have described the use
of fluorescence
Le., aflatoxins, citrinin,
techniques to detect and quantify selected mycotoxins,
ochratoxin A, and zearaleurone. Traditional detection methods for these mycotoxins have been either thin layer chromatography (TLC) or high-performance liquid
chromatography (HPLC) (70). The fluorescence characteristics
of aflatoxins (BI ,
B2,G1,andG2).citrinin,ochratoxin,
andzearaleuroneare listed in Table 3
(70). Because of the tediousness of the process involved
in TLC and HPLC,
investigation is being carried out by one of the coauthors (Panigrahi) to develop
a fast, nondestructive method for detection and/or quantification of mycotoxins.
He is using a personal computer-based
fiber optic spectroscope to investigate
fluorescence characteristics of wheat samples infected with mycotoxins and to
develop techniques to quantify the amount of mycotoxins present. Similarly, Lil1 ) studied the relationship between aflatoxin concentration and moislehoj et al. (7
ture at harvest and bright greenish-yellow fluorescenceof different corn hybrids.
Panigrahi et al. (72) investigated the autofluorescence characteristics
of edible beans for rapid, nondestructive evaluation. There are many types
of edible
beans: navy, pinto, red, black, etc. As with all other food and agricultural products, cracks or splitsin edible beans are undesirable andneed to be detected and/
or quantified. The conventional approach of crack detection in navy beans in-
Table 3 FluorescenceCharacteristics of Selected
Mycotoxins
Toxins
Aflatoxins
Citrinin and ochratoxin A
Zearalenone
Source: Adapted from Ref. 70
Excitation
wavelength
Emission
wavelength
340-380
254
440
400-700
308
400-700
124
volves dying them with a red dye and then counting the
of visible
numberdefects/
cracks. Use of red dye is justified because this enhances cracks and defects for
easy identification. The problem with this approach is that dyed beans are not
of
generally used for human consumption. It was found that the endosperms
the bean autofluoresce under UV light
(72). The fluorescence imaging system
(72) included a long-UV lamp for illumination.
developed for crack identification
CCD
The autofluorescenceof the beans was imaged using a low light-sensitive
camera. The acquired images were processed to quantify and assess the extent
of defect. Figure 11 is a typical image of a bean showing the bright defective/
cracked region that autofluoresces.
Autofluorescence has been used for quality characterization of fish meat,
to separate fishbone from meat, to quantify lean in animal carcasses, to detect
parasites in meat products, etc. (73-80). When excited at 340 nm, the fishbone
of a number of fishes (sod, whiting, haddock, plaice, flounder, and sole, etc.)
360 nm. The
fluoresces at 390 nm, but fish meat has a fluorescence peak at
intensity of fluorescence from the bones is also much higher than that from the
fish meat (Fig. 12) (81). These differences have been used to evaluate the quality
of fish fillets. Commercial fishbone detectorsare available that employ autofluorescence. The fluorescence intensity from fishbone decreases with the thickness
of the fish meat covering the bones. For example, when the flesh thickness of
0 to 2 mm, the relative fluorescence intensity from
cod and plaice increased from
the bones decreased almost linearly from 100% to about 10% (Fig. 13). Therefore, present technology is suitable for surface or near-surface detection, which
Fig. 11 Fluorescenceimage of an edible bean showing autofluorescing crackldefect

region.
125
100
75
50
25
0
350
450
400
500
Wavelength (nm)
Fig. 12 Fluorescence emission spectra of fishbone and fish meat when excited at
340
nm. (Ref. 81.) (Reprinted with permission from Jourrzul of Food Protecfion. 0 Interna-
tional Association for Food Protection.)

is not a serious limitation for the tish tillet industry. In general, the fluorescence
characteristics of fishbone are also valid for porcine, bovine, and chicken bones
and for cartilage and connective tissues(76,77,82). Therefore, techniques similar
to those used for fishbone detection may also be useful in meat processing.
In muscle foods, extensive and complex lipid-protein interactions generally
result from the peroxidationof unsaturated fatty acids. These interactions,in turn,
contribute to undesirable chemical changes such as insolubilization of proteins
and toughening of meat (83-85). The method commonly used to evaluate lipid
oxidation in meats is to measure the thiobarbituaric acid (TBA) value and perox-
0.5
1.5
Thickness (mm)
Fig. 13 Decrease in relative fluorescence intensity from fishbones with flesh thickness.
(Ref. 8 I .) (Reprinted with permission from Journal of Food P rorecrion. 0 International
Association for Food Protection.)
126
ide value (PV). TBS and PV values do not reliably relate to lipid oxidation in
meats (86), particularly in freeze-dried meat (86). Nakhost and Karel (88) used
fluorescence techniques to assess the effectiveness of antioxidantTBHQ (monot-butylhydroquinone), reduction of headspace oxygen, and compression on the
of oxidized beef
storage stability of freeze-dried meats. Fluorescence spectra
showed maximum excitation and emission at 350 and 440 nm, respectively. The
nonoxidized beef showed three peaks in the excitation spectrum at 308, 3 18, and
350 nm and a broad peak in the emission spectrum at about 475 nm. The use of
in beef samples,
antioxidant (100 ppm TBHQ) greatly decreased the fluorescence
and the fluorescence spectra of oxidized beef showed maximum excitation of
350 nm and emission excitation at 440 nm (89).
In the case of fruits and vegetables, as with
DLE, trends in chlorophyll
contentwerefollowedusingwhat
is knownaschlorophyllfluorescence.For
apples (90) and broccoli (91), fluorescence parameters were found suitable for
evaluating quality and maturity. Thus, use of chlorophyll fluorescence may be
suitable for maturity sorting of chlorophyll-containing fruits and vegetables on
commercial packaging lines (90). Woolf and Lasing (92) studied
avocado fruit
skin fluorescence following hot water treatments and pretreatments used as possible postharvest disinfection techniques. They reported that chlorophyll
fluorescence clearly reflected the effects of heat on the photosynthetic systems in
avocado fruit but did not detect the alleviation of heat damage by pretreatments.
Fluorescence spectra of fruit juices have also been proven useful quality
evaluation criteria (93,94). Seiden et al. (93) correlated the fluorescence spectra
of correctly
with soluble solids in apple juice and two apple cultivars as a means
to clasclassifying them. Similar fluorescence chemometrics were also suitable
sify and predict quality and process parametersof white sugar solutions and thick
juice samples of sugar beets.
B. InducedFluorescenceTechniques
In induced (secondary) fluorescence methods, food products are made
to fluoresce
by adding fluorochromes or by producing specific fluorescent reaction products
in the tissue or by applying fluorescent-labeled biological molecules (lectins antibodies) with specific binding affinity for the given food constituent. Table 4 lists
commonly used fluorochromes for food applications. The useof secondary fluorescence characteristics has been reported for many applications including the
use of calcoflour to detect glucans in malting barley, seed fixation system, use
of fluoroscamine to detect sprouting, and morphological and microchemical analysis of cereals (95,110). Fulcher et al. (96) provided a comprehensive overview
of fluorescence microscopy and its application to characterize nonstarchy carbohydrates and for routine analysis of products and processing conditions in industrial laboratories. Barnes and Fulcher (97) further reported about measuring fat
pplications
127
Table 4 CommonlyUsedFluorochromesfor
Food Applications
Fluorochrome
Safranin
Aniline blue
Congo red
Diphenylborinic acid
Acid fuchsin
Acridine orange
Acriflavine
Ruse Bengale
Calcoflour white
Auramine
Starch
Cellwalls
P-Glucans
Flavonoids
Storage protein
Microbiology, bacteria
E. coli
Live and dead yeast

Fungal element in food
Bacteria in food
Solrrcy: Refs. I I I , 112. (Adapted with permission.)
usinginducedfluorescence.According
to theirstudy(97),whenNileRed
(a
fluorescence dye) was added, aqueous suspension of wheat germ showed strong
fluorescence (Fig. 14). The absence of Nile Red, however showed almost no
fluorescence (Fig. 14) (97). The defatted powdered wheat germ (with Nile Red)
showed significant less fluorescence than normal wheat germ (Fig. 14) (97). This
implied the effect of fat content in emitting fluorescence. This postulation was
verified from the findings of Barnes and Fulcher (97) (Fig. 15). Figure 15 shows
strong dependency of fluorescence intensity on fat content of wheat flour mill
(measured by soxhlet extraction). Useof Nile Red for fluorescence measurement
of lipid and oils can be atool for quality control vegetable oil (97). The different
oils/lipids with dissolved Nile Red showed different emission wavelengths and
fluorescence intensities (Table5 ) (97). Tryglycerides (with Nile Red) were found
to have higher emission wavelengths than those
of polar monoglycerides and
unesterified fatty acids (with Nile Red) (97) (Table 5 ) . Fulcher et al. (69) examined the aleurone layer in barley seeds using fluorescence microscopy.
Fluorescence has also been used with cytometry for rapid analysis of food
microorganisms. The fundamentalprinciple of flowcytometry is asfollows:
a stationeryoptical
The sample to be examined is allowedtoflowthrough
sensing system in the form of continuous stream (using a flow cell). The optical
sensing system consists of a light source (excitation source) and a detector. The
of the
detector detects the optical signal representing the specific characterization
microorganism present in the sample (98). Maximum
flow rates up to a 0.5
mL/min and cell countratesupto
1O,OOO/s usingflowcytometryhavebeen
reported by Pinder and Godfrey (98). Fluorescent labeling technique, along with
flow cytometry, has shown potential for many structural and functional parame-
128
550
450
Emission Wavelength (nm)

Fig. 14 Fluorescence emission spectrum (arbitrary units) for an aqueous suspension of
powdered wheat germ (A) and defatted wheat germ (B) with Nile Red and without Nile
Red (C). (Ref. 97.) (Reprinted with permission.)
.-E
ln
c
3 .
10
Fat Content (%)
Fig. 15 Fluorescence intensity (arbitrary units) vs. fat content of mill streams from a
commercial wheat flour mill. (Ref. 97.) (Reprinted with permission.)
129
Table 5 Fluorescence Characteristics of Nile Red Dissolved in Lipids and

Commercial Vegetable Oils
Fluorescence Emission
ensity'
(nm) max.Lipidhegetable
(nm) max. oil
Excitation
Mono-olein
Oleic acid
Triolein
Monolinolein
Linoleic acid
Trilinolein
Corn oil
Olive oil
Palm oil
508
522
SO4
526
520
507
505
SO6
515
600
596
567
600
600
589
519
58 1
59 1
48
65
27 1
46
63
121
200
20 1
160
Relative. arbitrary units.

Source: Ref. 97. (Reprinted with permission.)
ters for food products. The structural and functional parameters are cell size, cell
shape, DNA and RNA content, surface sugars, total proteins, enzymes activity,
intracellular pH, DNA synthesis, etc. (98). Fluorescent flow cytometry has been
used to identify bacteriaby using light scattering and nucleic acid staining. Identification of food pathogens including Salmonella and Campylobacter bacteria has
been reported by immunofluorescent labeling (98). Other applications that have
usedflowcytometryinclude
analysis of fruit preparation, milk products, and
brewing yeast (98,99). Shelf life predictionof salads and fruit juice hasalso been
performed using flow cytometry (98).
Fluorescence cytometry, fluorescence microscopy, and other fluorescence
spectroscopic techniques have been presented
in more detail in other standard
sources (100,101).
VIII. SUMMARY
DLE and fluorescence are valuable techniques for evaluatinga variety of quality
factors.Allfruits,vegetables,andplantmaterialsundergoingphotosynthesis
probably produce DLE. However, the intensity and duration of the emitted light
vary widely, depending upon many factors. Because
of the strong dependence
of DLE on chlorophyll content, variationin DLE can be expected among different
varieties of the same product. Similarly, not all materials exhibit fluorescence,
and inducing fluorescence should be done carefully so as not to add agents that
would render foods unsafe. Fluorescence is also strongly dependent on environ-
130
mental and measurement conditions. Therefore, quality evaluation based on DLE

and fluorescence measurements requires careful selection of measurement criteria. It is necessary to obtain precise measurement protocols for a particular product and/orqualityfactorunderasetcondition,andmeasurementconditions
should be carefully validated to establish standard measuring criteria.
Somemeasurementrequirements ofboth DLE-andfluorescence-based
quality evaluation systems may not match the high throughput values
of more
traditional spectrophotometric or calorimetricunits. The operating speed of newer
computer visioninspectionsystems is alsomuchhigher.However,DLEand
in situations where product quality
fluorescence measurements can play a key role
in physiological charis not readily apparent but based on very subtle differences
acteristics and in cases where accuracy rather than speedof operation is of prime
consideration. Fluorescence measurements, for example, are very widely used
for microscopic measurements that are not possible with other measurement techniques. New advances such as photoluminography, which is imaging of chlorophyll-containing products using DLE ( I 02,103). and fluorescence imaging will
make DLE and fluorescence applications more commonplace. With continued
research and innovation, quality evaluation systems based on DLE and fluorescence measurements will find wider applications.
REFERENCES
I.
2.
3.
4.
5.
6.
7.
8.
9.
IO.
1I.
NA Eskin, HM Henderson, RJ Townsend. Biochemistry of Foods. NewYork: Academic Press, I97 I , p 46.
BL Strehler, W Arnold. Light production by green plants. J Gen Physiol 342309,
1951.
GG Dull. Nondestructive evaluationof quality of stored fruits and vegetables. Food
Techno1 40(5): 106- 1 IO, 1986.
WR Forbus, SD Senter, RL Wilson. Measurement of tomato maturity by delayed
light emission. J Food Sci 50(3):750, 1985.
H Wain. The Story of Fluorescence. Middlefield, CT: RayTech Company, 1965.
GGGuilbalt.PracticalFluorescence:Theory,Methods,andTechniques.New
York: Marcel Dekker, Inc., 1973, pp 1-45.
S Itoh and N Murata. Correlation between delayed light emission and fluorescence
of chlorophyll A in system I1 particles derived from spinach chloroplasts. PhotochemPhotobiol18(3):209,1973.
G Tollin, M Calvin. The luminescence of chlorophyll-containing plant material.
Proc Natl Acad Sci 43:895, 1957.
G Tollin, E Fujimori, M Calvin. Delayed light emission in green plant materials:
temperature-dependence and quantum yields. Proc Natl Acad Sci 44:1035, 1958.
DF Bradley, M Calvin. The effect of thioctic acid on the quantum efficient of the
Hill reaction in intermittent light. Proc Natl Acad Sci
41563. 1955.
PB Sogo, NG Pon, M Calvin. Photospin resonance in chlorophyll-containing plant
material. Proc Natl Acad Sci 43:387, 1957.
131
12. W Arnold, JB Davidson. The identity of the fluorescent and DLE spectra in Chlorelln. J Gen Physiol 37:677, 1954.
primary process in photosynthesis.I. Photo13. WE Arthur, BL Strehler. Studies on the
70(2):507,
synthesisluminescence:multiplereactants.ArchBiochemBiophys
1957.
14. W Bertsch. J West, R Hill. Delayed light studies on photosynthetic energy conversion. 11. Effect of electron acceptors and phosphorylation cofactors on the millisecond emission from chloroplasts. Biochim Biophys Acta 172(3):525, 1969.
15. P Sweetser, CW Todd, RT Hersch. Effect of photosynthesis inhibitors on light reemission in photosynthesis. Biochim Biophys Acta 51(3):509,1961.
16. S Itoh, N Murata, A Takamiya. Studies on the delayed light emission in spinach
chloroplasts. I. Nature of two phases in development of the millisecond delayed
light emission during intermittent illumination. Biochim Biophys Acta
245(I): 109,
197 I .
17. BC Mayne. The effectof inhibitors and uncouplers of photosynthetic phosphorylation on the DLE of chloroplasts. Photochem Photobiol 6:189, 1967.
18. CA Wraight, AR Crofts. Delayed fluorescence and the high-energy state of chloroplasts. Eur J Biochem 19(3):386, 1981.
19. RM Groves. A pulsed laser system for the study of millisecond delayed fluorescence from plants. Photochem Photobiol 27(4):491, 1978.
20. Y Chuma, K Nakaji. Optical properties of fruits and vegetables to serve the automatic selection within the packaging house line.
IV. Delayed light emission as a
means of automatic selection of tomatoes. J SOC Agric Mach
(Japan) 38(2):217,
1976.
21. K Nakaji, Y Sigiura, Y Chuma. Delayed light emission of green tea
as a means
of quality evaluation. I. Delayed light emission of fresh tea leaves. J SOC Agric
Mach (Japan) 39(4):483, 1978.
22. Y Chuma, K Sein, S Kawano,KNakaji.Delayedlightemission
as a means of
automatic selection of Satsuma oranges. Trans Am SOC Agric Eng
20(5):996. 1977.
23. Y Chuma, K Nakaji, M Ohura. Maturity evaluation of bananas by delayed light
emission. Trans Am SOC Agric Eng 23(4):1043, 1980.
24. FC Jacob, RJ Romami, CM Sprock. Fruit sorting by delayed light emission. Trans
Am SOC Agric Eng8(I): 18, 1965.
25. P Jursinic, Govindjee. Temperature dependence of delayed light emission in the
6-340microsecond range after a single flash in chloroplasts. Photochem Photobiol
20(6):617, 1977.
26. JA Abbott, DR Massie. Delayed light emission for early detection of chilling in
cucumber and bell pepper fruit. J Am SOC Hort Sci
110(1):42, 1985.
27. WB Terzaghi, DC Fork, JA Berry, CB Field. Low and high temperature limits to
PSII. A survey using trans-parinaric acid, delayed light emission, Fo
andchlorophyll
fluorescence. Plant Physiol 9l(4): 14941500, 1989.
28. AZ Watada. KH Norris, JT Worthington, DR Massie. Estimation of chlorophyll
and carotenoic contents of whole tomato by light absorbance techniques. J Food
Sci 41:329, 1976.
29. WR Forbus, Jr.. SD Senter. HT Chan,Jr. Measurement of papaya maturity by delayed light emission. J Food Sci 52(2):356, 1987.
30. GL Graf, GE Rehkugler, WFE Millier, JA Throop. Automatic detection of surface
132
flaws on apples using digital image processing. Am SOC Agric Eng Paper No. 813537, 1981.
31. N Sarkar, RR Wolfe. Feature extraction techniques for sorting tomatoes
bycomputer vision. Trans Am SOC Agric Eng 28(3):970, 1985.
32. RR Wolfe, M Swaminathan. Determining orientation and shape of bell peppers by
machine vision. Am SOC Agric Eng Paper No. 86-3045, 1986.
33. S Gunasekaran, MR Paulsen.Automatic,nondestructivedetectionofcornkernel
defects. In: WJ McGonnagle, ed. International Advances in Non-destructive Testing, Vol. 12. New York: Gordon and Breach Science Publishers, 1986,p 95.
34. S Gunasekaran, TM Cooper, AG Berlage. Quality evaluation of grains byimage
processing. Proceedings International Symposium Agricultural Mechanization International Cooperation High Technology Era, 1986,p 327.
35. S Gunasekaran, TM Cooper,AGBerlage.Imageprocessingforstresscracksin
corn kernels. Trans Am SOC Agric Eng 30(1):266, 1987.
36. TMatshuisa,AHosokawa.Possibilitiesofcheckingcracksofbrownnceusing
illumination by oblique ray and image data processing system. J SOC Agric Mach
42(4):515, 1981.
37. S Gunasekaram, TM Cooper, AG Berlage. Evaluating quality factors of corn and
soybeans using a computer vision system. Trans Am SOC Agric Eng3 l(4): 126f41271, 1988.
38. Y Chuma, K Nakaji. Delayed light emission as a means of color sorting of plant
products. In: P Linko, Y Malkki, J Olkku, J Larinkari, eds. Food Process Engineering, Vol. I. London: Applied Science Publishers, 1979,p 314.
39. JA Abbott, DT Krizek, P Semeniuk, HE Moline, RM Mirecki. Refreshed delayed
light emission and fluorescence for detecting pretreatment effects on chilling injury
in coleus. J Am SOC Hort Sci 112(3):560, 1987.
40. RM Smillie. The useful chloroplast: a new approach for investigating chilling stress
in plants. In: J Lyons, JK Raison, D Graham, D, eds. Low Temperature Stress
in
Crop Plants. New York: Academic Press, 1979.
41. JA Abbott, AR Miller, TA Campbell. Detection of mechanical injury and physiological breakdown of cucumbers using delayed light emission. J Am SOC Hort Sci
116(1):52-57, 1991.
42. Agriculturalresearch,newdetectorforhiddeninjuries
in fruit.AgricRes 34(1):
15, 1986.
43. ST Seo, GJ Farias, EJ Harris. Oriental fruit fly: ripening of fruit and its effect on
the index of infestation of Hawaiian papayas. J Econ Entomol 75:173, 1982.
44. WR Forbus,Jr.,HTChan,Jr.Delayedlightemissionasameansofpredicting
papaya susceptibility to fruit fly infestation. J Am SOC Hort Sci 114(3):521-525,
1999.
45. WR Forbus, SD Senter. Delayed light emission as an indicator of cataloupe maturity. J Food Sci 54(4): 1094-1095, 1989.
46. JARobertson,FIMeredith,WRForbus,BGLyon.Relationshipofqualitycharacteristicsofpeaches(Cv.Loring)tomaturity.
J FoodSci 57(6):1401-1404,
1992.
47. JA Robertson, FI Meredith, WR Forbus. Changes in quality characteristics during
peach (Cv. Majestic) maturation. J Food Quality 14(3):197-207, 1991.
133
48. WR Forbus, GG Dull. Delayed light emission as an indicator of peach maturity.

J Food Sci 55(6):1581-1584, 1990.
49. WR Forbus, JA Payne, SD Senter. Nondestructive evaluation of Japanese persimmon maturity by delayed light emission. J Food Sci 56(1):985-988, 1991.
50. FI Meredith, SD Senter, WR Forbus, JA Robertson, WR Okie. Postharvest quality
and sensory attributes of Byrongold and Rubysweet plums. J Food Quality
15(3):199-209,1992.
by delayed
51. WR Forbus, GG Dull, D Smittle. Measuring netted muskmelon maturity
light emission. J Food Sci 56(1):981-984, 1991.
52. WR Forbus, GG Dull, DA Smittle. Nondestructive measurement of canary melon
maturity by delayed light emission. J Food Quality 15(2): 119-127, 1992.
53. WR Forbus, Jr., GA Hardigree, JH Adams. Experimental delayed lights emission
meter for horticultural crops. Agricultural Research Service Bulletin No. 41, 1985.
54. F Grum, RJ Bechever. Optical Radiation Measurements, Vol. I , Radiometry. New
York:AcademicPress,1979.
55. F Grum, CJ Bartleson, eds. Optical Radiation Measurements, Vol. 2, Color Measurement. NewYork:AcademicPress,1980.
56. KS Mielenz, ed. Optical Radiation Measurements, Vol. 3, Measurement of Photoluminescence. New York: Academic Press, 1982.
57. S Gunasekaran, MR Paulsen, GC Shove. A laser optical method for detecting corn
kernel defects. Trans Am SOC Agric Eng 29(1):294, 1986.
58. Z Sun, P Chen. Distributionof light transmitted through agricultural products. Am
SOC Agric Eng Paper No. 83-3089, 1983.
59. GS Birth. Electromagnetic radiation: optical. In: Instrumentation and Measurement
for Environmental Sciences. ZA Henry, ed. St. Joseph, MI: American Society of
Agricultural Engineers, Supplementary Publication No. SP-0375, 1975.
60. DR Massie, KH Norris. A high-intensity spectrophotometer interfaced witha computer for food quality measurement. Trans Am SOC Agric Eng 18( I ) : 173, 1975.
61. JM Henderson, BM Shawver. Singulation-a problem in design. J Eng Ind 2:101,
1973.
62. BM Shawver. JM Henderson. A compendium of singulation and associated devices.
Department of Mechanical Engineering Report, University of California, Davis,
CA,1982.
63 TH Burkhardt, RF Mrozek. An orienting and conveying device for sorting dried
prunes. Trans Am SOC Agric Eng 17(6):1173, 1974.
64 JJ Gaffney. Light reflectance of radishes as a basis for automatic grading. In: JJ
Gaffney, ed. Quality Detectionin Foods, Vol. 5 . St. Joseph, MI: American Society
of Agricultural Engineers, 1976, p 75.
65 MP Fogarty, CN Ho, IM Warner. Data handling in fluorescence spectrometry. In:
KD Mielenz, ed. Optical Radiation Measurements, Vol. 3, Measurement of Photoluminescence. New York: Academic Press, 1982,
p 249.
66 WF McClure, RP Rohrbach, LJ Kishmam, WE Ballinger. Design of a high-speed
fiber optic blueberry sorter. Trans Am SOC Agric Eng 18(3):487, 1975.
67. AG Story, GSV Raghavan. Sorting potatoes from stones and soil clodsby infrared
reflectance. Trans Am SOC Agric Eng 16(2):304, 1973.
68. CF Earp, CA Doherty,LW Rooney. Fluorescence microscopy of pericarp, aleurone
134
Panigrahi
and
Gunasekaran
layer,andendospermcellwallsofsorghumcultivars.CerealChem60(5):408410,1983.
69. RG Fulcher.GSelterfield,MMcCullyd,PWord.Observationsonthealeurone
layer 11. Fluorescence microscopy of the Aleurone sub-aleurone junction
as the
emphasis on possible B 1,3 glucan deposits in barley. Aust J Biol Sci 25:23-34,
1972.
70. GM Wood, PJ Mann. Detection of selected mycotoxins in foods by fluorescence.
In: Fluorescence Analysis in Foods. L Munck, ed. Longman Group, 1984, pp 158170.
71. EB Lillehoj, A Manwiller, TB Whitaker, MS Zuber. Hybrid difference in estimation
of preharvest occurrence of bright greenish-yellow fluorescence and aflatoxin in
corn. J Environ Quality 12(2):216-219, 1983.
72. S Panigrahi. H Jiao, J Lindsley. Computer vision methods for defect detection of
edible beans. Am SOC Agric Eng Paper No. 97-3130, 1997.
73. SAa Jenson, S. Reenberg, L. Munck. Fluorescence analysis in fish and meat technology. In: Fluorescence Analysis in Foods. L Munck, ed. Longman Group, 1989,
pp 182- 192.
74. SP Aubourg, I Medina. Quality differences assessment in canned sardine (Sardina
pilchardus) by fluorescencedetection. J AgricFoodChem45(9):3617-3621,
1997.
75. JJ Swatland. Relationship between the back-scatter of polarized light and the fibreoptic detection of connective tissue fluorescence in beef. J Sci Food Agric 75( I):
45-49,1997.
76. HJ Swatland, CJ Findlay. On-line probe prediction of beef toughness, correlating
sensory evaluation with fluorescence detection of connective tissue and dynamic
analysis of overall toughness. Food Quality Preference 8(3):233-239, 1997.
77. HJ Swatland, C Warkup, A Cuthbertson. Testing a UV fluorescence probe for beef
carcass connective tissue. Computers Electron Agric 9(3):255-267, 1993.
78. P Velinov, RG Cassens, ML Greaser, JD Fritz, ME Cassens. Evaluation of meat
products by fluorescence microscopy. J Muscle Foods 2(1):57-63, 1991.
79. SP Aubourg. Effect of pH on fluorescence formation related to fish deterioration.
Food Res Techno1 207(4):268-272, 1998.
80. SP Aubourg, CG Sotelo, R.Perez-Martin. Assessment of quality changes in frozen
sardine (Sardirza pilchardus) by fluorescence detection. J Am Oil Chem SOC 75(5):
575-580,1998.
8 I . HH Huss, P Sigsgaard, SAa Jensen. Fluorescence of fish bones. J Food Prot 48(5):
393-396,1985.
82. SAn Jensen, L Munck, P Sigsgaard, HH Huss. Method for quality control of products from fish, cattle, swine and poultry. U.S. Patent 4,631,413 (1986).
83. JD Love, AM Pearson. Lipid oxidation in meat and meat products: a review. J Am
Oil Chem SOC 48547-549, 1971.
84. Z Nakhost,MKarel.Measurementofoxidation-relatedchangesinproteinsof
freeze-dried meats. J Food Sci 49: 1 171- 1 173, 1984.
85. Z Nakhost, M Karel. Effect of salt, tripolyphosphate, and tertiary butylhydroquinone on myoglobin-lipid oxidation indicatorsin freeze-dried meats. J Food Sci 50:
1748-1749,1985.
135
86. SL Melton. Methodology for following lipid oxidationin muscle foods. Food Techno1 37:105-140,1983.
87. JR Chipault, JM Hawkins.Lipidoxidationinfreeze-driedmeats.
J AgricFood
Chem 19:495-499, 197 1.
88. Z Nakhost, M Karel. Fluorescence due to interactions of oxidizing lipids and proteins in meats. In: Fluorescence Analysis in Foods. L Munck, ed. Longman Group,
1989,pp193-206.
89. AR Kamarei, M Karel. Assessment of auto-oxidation in freeze-dried meats by a
fluorescence assay. J Food Sci 49: I517-1520, 1524, 1984.
90. J Song, W Deng,RM Beaudry, PR Armstrong. Changesin chlorophyll fluorescence
ofapplefruitduringmaturation.ripening,andsenescence.HortSci32(5):891896, 1997.
91. PMA Trivonen, JR DeEII. Differences in chlorophyll fluorescence and chlorophyll
content of broccoli associated with maturity and sampling section. Postharvest Biol
Technol14(1):61-64,1998.
92. AB Woolf, WA Lasing. Avocado fruit skin fluorescence following hot water treatments and pretreatments J Am Soc Hort Sci 12 1(1): 147- IS I , 1996.
93. P Seiden, R Bro, L Poll, L Munck. Exploring fluorescence spectra of apple juice
and their connection to quality parameters by chemometrics
J Agric Food Chem
44( 10):3202-3305, 1996.
94. L Norgaard. Classification and prediction
of quality and process parameters of thick
juice and beet sugar by fluorescence spectroscopy and chemometrics. Zucker Ind
120(1):970-981,1995.
9s. RG Fulcher. Fluorescence microscopy of cereals. Food Microstructure 75(
I): 167175.1982.
96. RG Fulcher, PJ Wood, SH Yiu. Insights into food carbohydrates through fluorescence microscopy. Food Technol (Jan): 101- 106, 1984.
97. PJ Barnes, RG Fulcher. Fluorometric measurement of fats. In: L Munck, ed. Fluorescence Analysis in Foods. New York: John Wiley and Sons, Inc., 1989.
of food micro98. AC Pinder, S Godfrey. Fluorescence cytometry for the rapid analysis
organisms. In: Food Process Monitoring Systems. AC Pinder, G Godfrey, eds.
New
York: Blackie Academic and Professional, 1993.
99. L Jespersen. Use of fluorescence staining and flow cytometry for analysis of brewingyeasts (Snccharornyces cerevisiae). DissertationAbstractsInt57(
I ) : 1 IO,
1996.
100. OS Wolfbeis, ed. Fluorescence SpectroscopyNew Methods and Applications. New
York:Springer-Verlag. 1993.
101. J Slavik.FluorescenceMicroscopyandFluorescentProbes.NewYork:Plenum
Press,1996.
102. LO Bjorn, AS Forsberg. Imaging by delayed light emission (photoluminography)
as a method for detecting damage to the photosynthetic system tested on leaves of
Hibiscus,Oxalistetraphylla,tobacco,Fagussilvatica,Polypodiumvulgareand
beans. Physiol Plants 47(4):215-222, 1979.
103. ESundborn,LOBjorn.Phytoluminography:imagingplants
by delayedlight
emission (for the study of pathological disturbances). Physiol Plants 40( 1):39-41,
1977.
136
104. JR Lakowicz. Principles of Fluorescence Spectroscopy.

1 05.
106.
107.
108.
109.
1 10.
111.
112.
113.
New York: Plenum Publishing, 1993, p 48.

JE Kaufman, ed. IES Lighting Hand
Book. New York: Illuminating Engineering
Society,1972.
L Godfrey. Choosing a detector for your light sensing application. Laser Focus
World (March):I09-119, 1995.
K Kaufman. Detectors cover the spectrum of instrument applications. Laser Focus
World. (Feb.):99-105, 1994.
Product Catalog 1999-2000. New Focus, Inc., Santa Clara, CA, 1999, pp 56-62.
GM Williams, MM Blooke. How to capture low light level images without intensitiers. Laser Focus World (September): 1995.
L Munck. Practical experiences in development of fluorescence analyses in an applied food research laboratory. In: Fluorescence Analysis in Foods. L Munck, ed.
New York: Longman Group/John Wiley and Sons, 1989, pp 3-29.
RG Fulcher, DW Irving, A de Francisco. Fluorescence microscopy: applications
in food analysis. In: Fluorescence Analysis in Foods. L Munck, ed. New York:
Longman Group/John Wiley and Sons, 1989, pp 85-86.
P Sigsgaard. Fluorescence microscopy techniques in the practice of food microbiology.In:FluorescenceAnalysisinFoods.LMunck,ed.NewYork:Longman
GrouplJohn Wiley and Sons, 1989, pp 131-139.
GG Guilbault. Principles of fluorescence spectroscopy
in the array of food products.
In: Fluorescence Analysis in Foods. L Munck, ed. New York: Longman Group/
John Wiley and Sons, 1989, pp 34-46.
X-Ray Imaging for Classifying Food

Products Based on Internal Defects
Ernest W. Tollner
The University of Georgia, Athens, Georgia
Muhammad Afzal Shahin
Canadian Grain Commission, Winnipeg, Manitoba, Canada
1.
INTRODUCTION
Increased consumer awareness of product quality is making the marketing of

food products, ranging from fresh produce to processed foods, very competitive
in that maintaining high quality is essential to economic survival in the business.
Products may be damaged during harvest and postharvest operations (external
defects). Products may also have some internal defects ranging from foreign materials (e.g., metal fragments, bone fragments), watercore (e.g., apples), andvaria direct adverse effect
ous neckrots (e.g., onions). These quality defects have
on the market value of the products. Consumers desire consistent, high-quality
products, thus, removal of defective units from the batch becomes absolutely
necessaryforcustomersatisfaction-akeytosuccess
in anybusiness.X-ray
techniques are widely used commercially for detecting bone fragments and other
foreign objects in processed foods. Commercial application of x-ray-based techniques for inspecting fresh foods will likely be realized
in the near future.
as a
At present, visual inspection, though very subjective, is being used
means to determine external food quality. Some researchers have also reported
the use of computer vision systems, and these techniques are beginning to see
commercial implementation. However, product classification based on internal
quality is almost nonexistent commercially. Internal qualityis typically estimated
a product lot viaa destrucby visual inspection of product randomly picked from
137
Tollner and Shahin
I 38
tive test. A major limitation of this method, besides being destructive,

is that
internalqualityestimation is basedonrandomsamples.Henceproductswith
internal defects may go undetected. In the case of fruits and vegetables, storage
of products with internal defects may result in spoilage of surrounding healthy
produce during storage.
Due to a shortage of labor force willing to accept short-term employment
during the harvest season, the fruit and vegetable industry is considering the use
of mechanical harvesters as the obvious solution to overcome the problem. The
use of mechanical harvesters, which may cause more bruise damage to the fruit,
is anticipated to require a thorough inspection of individual fruits. Again, due to
lack of experienced workers for food quality evaluation, the need for automatic
sensing techniques has increased. It is also desirable that the technique is nonso that even internal defects can be
destructive and can be performed on-line
detected by examining each of the products being processed or shipped.
Internal defects may cause downgrading or even rejection of the whole lot
for fresh market-a big loss to the growers. For processors, foreign object detection by consumers can be economically devastating, resulting in the need to remove contaminated product. For packers, the storage cost of fruits with internal
defect decreases their profit margin. Moreover, internally damaged units pose a
threat of tissue breakdown during storage that can lead to the loss of the whole
batch. Hence, detection of internal quality is essential in order to make a decision
,"\
Pterpaf,
Economic
Inputs
,Qua ~ t y
I
/I
\\
I
I
Alternatwe
Store fol
ater supp y
4
p$
ar et
Fig. 1 Schematicdiagram of the overalldecisionsupport system (DSS) showingthe

fruit supply decision based on quality and economic inputs. Fruit quality decision based
on internal defects was of immediate interest.
X-Ray Imaging
139
whether the product should be marketed immediately, stored for supply later on
to maximize profit, used for making alternative products, or discarded.
Automation of the quality evaluation process could generate significant
economic benefits to growers, packers, and processors. Our overarching view of
the decision support system for food systems is shown in Fig. 1. The component
enclosed by the dotted line in Fig. I , concerned with internal quality, is the primary subject of this chapter.
Various energy-based imaging techniques have been
tried in the past for
the detection of internal defects with varying degrees of success. The most appropriate image analysis technique that can relate image features to product quality
is, in many cases, yet to be identified. An emphasis should be placed on the
development of better image analysis and decision-making techniques for the
widest range of inspection applications. The goal of this chapter is to emphasize
x-ray techniques in the food industry and present some fairly promising case
study results and review relevant applications.
II. FUNDAMENTALS OF X-RAYIMAGING

A.
Image Acquisition
X-ray imaging is a radiographic method that provides a cross-sectional view of

an object's interior. Specification of an x-ray system involves selection
of an
energy level, residence time or signal intensity, product orientation, and scan type
(computer tomography, CT, or line scan, LS). A typical x-ray imaging system
consists of an x-ray source, a detector array, and mechanical systems to rotate
and move the components relative to the source and detectors (Fig. 2 ) . When xrays pass through a material, they are partly absorbed along the way by the
test
specimen. The intensity is reduced according to an absorption coefficient p(x,y),
which depends on the local elemental composition and density of the test specimen.Acomputer is used toreconstructanimage
of across-sectionalplane
through an object. With x-ray CT imaging, it is possible to virtually slice open
the test object, examine its internal features, and identify internal defects
that
may exist. The process of x-ray absorption is described by Beer's law given as:
where
I
I,,
=
=
x-ray photon intensity striking the detectors (photons/s)

initial x-ray photon intensity (photons/s)
x-ray absorption coefficient (m")
length of the projection transact through the test object (m)
1.Line scan voxei

CT voxel
45 4 r e e s
180 degrees
315 degrees
90 degrees
es
360 degrees
Fig. 2 Three-dimensional schematic (a) of the x-ray imaging system showing two modes of operation: line scan (LS) and CT scan. In the LS
mode, the object moves across stationary source-detector assembly. The LS voxel is indicated along with a corresponding LS pixel in (a). In the
C T mode the source-detector assembly rotates around the stationary object as indicated in (b). Fig. (b) is an end view of Fig. (a), rotated the
indicated degrees. The CT voxel is indicated, as is the corresponding pixel in (a).
X-Ray Imaging
141
The measurement of the x-ray absorption coefficientp (or some parameter related
to p) is of paramount interest in x-ray applications. The image is in effect a map
of p. X-ray absorption in food products is partitioned into solids pg and water
p wcomponents, as shown below:
where
vu, = x-ray absorption coefficient of water at the specified energy level (m")
p, = x-ray absorption coefficient of the solids at the specified energy level
pw=
ps =
pSgw=
p,$\ =
(m")
mass of water per unit volume of sample (kg/m3)
dry bulk density of the sample (kg/m')
density of water (assumed constant, kg/m3)
particle density of dry solids (assumed constant, kg/m')
Note that p\+/pbg\\

is equivalent to the volumetric moisture content and p,/pSFIis
equivalent to the volumetric solids content. In most applications involving food
products, the water term dominates and the solids dry density term may be neglected. Thus, one may infer that CT and LS map water content of the product.
on
The x-ray absorption coefficients for the dry solids and water depend
the energy level of the x-rays (1). Thus, in theory, one could scan a sample with
x-rays of different energy. Knowing the x-ray absorption coefficient for each
phase at each energy level, one could then solve for the water density and solids
density. (Detailed discussions followin a later section.) X-rays with energy levels
in the 120 kV or
in the 5- 15 kV range are much more easily absorbed than
higher range. Thus, absorption coefficients are inversely proportional
to energy
level. Therefore, lower energy levels are more appropriate for small product inspection. CT scanners operate at 120 kV or higher due to the need to penetrate
large samples. The available energy level with most x-ray machinesis fixed. One
would generally select the lowest usable energy level.
X-ray spectral analyzers enable oneto quantify x-ray intensity as a function
of x-ray energy level. In a study involving potential foreign objects in processed
foods, Morita et al. (2), using 25 and 50 kV x-ray sources, observed noteworthy
differences in the absorption of acrylic materials. This bodes well for foreign
material identification and removal from processed foods.
The two scanning modes for generating
x-ray images, CT and LS, are schematically illustrated in Fig. 2. In the CT mode, a series of projections around the
stationary test object is obtained to reconstruct the image. The reconstructed image is a quantitative map of p(x,y) at each point in the scanned plane. The CT
voxel is acube(typically, 2 mm X 2 mm X 2 mm)definedbycollimation
Shahin
142
and
Tollner
thickness and number of projections. CT image pixel is shown over a scaled 2

mm X 2 mm region at intensity corresponding to the average absorption value
for the CT voxel. CT scans provide an accurate view of the interior of the object
and work well for off-line research applications. Filtered back-projection using
in case of
an ideal inverse filter produces a good-quality, high-contrast image
low noise projection data. Noisy projection data, however, result in poor quality
image (low signal-to-noise ratio) as the inverse filter tends to enhance the highfrequency details (noise). CT scanning has become well established asan inspection, evaluation, and analysis
tool in medical diagnosis and other research applications. As one may discern from the numbers of projections indicated in Fig. 2
required for a CT image, CT is primarily a research tool in nonmedical applications. However, modern CT scanners are now using electronic waveguides
(as
opposed to mechanically moving the x-ray source and detectors) to very rapidly
collect projection data. By using hardware approaches (as opposed to software),
image reconstruction can be very fast. Thus,CT may be a viable tool for selected
inspection applications in the near future.
In the LS mode, projection data are collected as the object moves between
the x-ray source and detector hardware. The reconstructed image in the LS mode
is a quantitative map of p(x,y) along numerous collinear paths. The LS voxel is
a rectangle defined by the collimator resolution and depth of the test object. The
LS pixel is displayed such that the typically 2 mm X 2 mm pixel dimension is
shown at a brightness level correspondingto the average relativex-ray absorption
of the LS voxel. Line scan imaging provides relatively less accurate image of
the object interior (because of the averaging over the projection line). However,
because of the speed of operation, the LS mode is more appropriate for on-line
inspection applications than the CT mode. Luggage inspection at airports is one
familiar example of line scan imaging applications. Line scanning offers the potential to automate the process of produce classification based on quality. Therefore, x-ray line scan imaging is the most suitable for on-line inspection. Line
in potatoes
scanners have been commercially used for hollow heart detection
for nearly I O years (3). The x-ray line scanners are becoming commercialized
particularly for foreign object detection in food products (4), and increased commercial use is expected in the near future.
X-ray image intensifiers enable one, in effect, to magnify small defects.
Approaches involving intensifiers are proving useful for inspecting small grains
and nuts for insect larvae presence (5). This technology is currently a research
tool. Additional details of x-ray absorption for food quality evaluation applications are described by Garrett and Lenker (6).
B. ProductOrientation
Product orientation has a direct effect on the feature characteristics. Orientation
relative to the scanned plane affects both shape and brightness of features. These
X-Ray
143
features bear on the Gestalt approach

to perceptual organization (7). For example,
watercore in apples does not always occur as a simple blob. Instead, it occurs
as a constellation of blobs and is probably a three-dimensional feature. X-ray
imagesreducethree-dimensionalfeaturestotwo-dimensionalfeatures(unless
one does exhaustive serial CT scans and constructs a three-dimensional representation).Furthercomplicatingtheissue,theexpertsourcesmakeobservations
based on two-dimensional cuts, which are usually somewhat random.
The experts
loss of
do not articulate the nature of the visualized blob relationships. This
intelligence hinders the automation process.
Because LS integrates along zones through the product, one can influence
the degreeof absorption by causing muchof the featureto be containedby voxels.
One may also orient the product such that some of the feature occursin as many
voxels as possible. Preliminary studies are generally sufficient for this determination. For watercore and bruise determination in apples it was observed that scanning along the stem-calyx axis was clearly preferable to other orientations. The
of the bruised tissue with
stem-calyx orientation enables minimal interference
sound tissue since most bruises occur on the fruit periphery. There may be an
advantage to having some constancy in physiological features (e.g., the core of
an apple usually shows as a star-shaped void when line scanning apples along
thestem-calyxaxis). CT scanningplanesgenerallyprovidemostinformation
when the feature of interest occupies as many pixels as possible. For example,
the extent of watercore is readily observed by orienting the fruit such
that the
stem-calyx axis is normal to the scanned plane and positioned such that the scan
plane intersects the maximum fruit diameter.
C. Image Noise
The image analysis processis essentially one of extracting desired image features
(if present) from the vast arrayof possible undesirable image features. The undesirable features arise from a variety of sources: the instrument, image artifacts,
image features not related to the quality attribute at hand, and lack of fit of the
decision engine algorithm and the distribution of the desired feature. Each proto decrease the
cessing step in an image processing protocol must be designed
into nondesirnondesirable features without transforming some desirable features
able features. At this point we consider instrument-induced noise and image artifacts. Other noise sources will be considered in the discussion of feature extraction and decision engine selection.
During the process of image acquisition, noise is introduced inadvertently
to theactualsignal,whichmustberemoved
or at least minimized to extract
features related to the quality of the product. Instrument-induced noise is in addition to noise associated to unrelated features of interest in an ideal x-ray image.
An x-ray image may be subject to instrument noise and interference from several
sources, including electrical sensor noise, channel errors, spatial sampling and
Shahin
144
and
Tollner
gray-level quantization noise, reconstruction errors, etc.

The use of a window
function such as a Hamming window (8) leads to a superior image reconstructed
from projections. The window function deemphasizes high-frequency components, which mostly represent noise (9).
The most common source of noise is counting statistics in the image detecof incident particles (photons, electrons, etc.). The
tors due to a small number
number of x-ray photons detected at each point in the projection set
is subject
to fluctuations due to counting statistics. The number of detected x-rays varies
in a Gaussian or normal distribution whose standard deviation is the square root
of the number counted (IO). The noise is, however, inversely proportional to the
square root of the number detected. Counting100 x-rays per detector, on average,
produces a noise whose standard deviation is 10%; while on average 10,000 xrays are needed to reduce the variation to 1 %. Noisy images may also occur due
to instability in the sourceor detector during the time required
to scan or digitize
an image. The pattern of this noise may be quite different from the essentially
Gaussian noise due to counting statistics. Modern systems are designed to give
warnings when lower than desired x-ray photons are detected.
Additional sources of noise are particularly important with
CT imaging.
Beam hardening describes the effect in which the lower energy softer x-rays
in a sample. As the lower
from a polychromatic source are preferentially absorbed
energy x-rays are absorbed, the attenuation coefficient
of the sample changes
independent of any actual change in composition or density. Beam hardening is
not a major problem in biological and medical applications. It can be a serious
problem in industrial applications where the composition and density within samof each image feature changes according
ples vary over a wide range. The contrast
to where it lies within the object with respect to the source. Diffraction and scattering of x-rays also affect the quality of images in a way similar to that of beam
hardening. Beam hardening and diffraction effects are minimized
by avoiding
metals and other high x-ray-absorbing materials.The avoidance of sharp boundaries in the scanned region controls these difficulties in CT applications. They
are not major problems in LS applications. Appropriate preventive maintenance
and frequent calibrations minimize these errors. A window function placed
in
the reconstruction algorithm minimizes artifacts due
to quantization error and
sensor noise (9).
The process of image reconstruction from projections amplifies the effect
of noisein CT images because the
filtering process suppresses the low frequencies
of errors in the
and enhances the high frequencies. Another important source
reconstruction of CT images is imprecise knowledge of the location of the center
of rotation or variation in that center due to imperfect mechanical mechanisms
( 1 1). When the objectis not exactly in the center of rotation, circular or U-shaped
artifacts appear in the reconstructed image. Usually, a large number of views and
to provide enough
enough detector positions along each projection set are used
X-Ray imaging
145
information for CT reconstruction. In the event that fewer projections in a set or

fewer views are used, the image has more artifacts and poorer resolution.
Shahin et al. ( 1 2) evaluated noise levels in a CT and LS application and
discussed an approach for designing filtering masks for minimizing the impact
of using
of image noise on accuracyof feature extraction. The approach consisted
correlation and variography for assessing image noise and noise patterns, where
or from waves in the case of water
patterns may arise from imaging artifacts
transport. Filter mask type and size may then be proposed. Bracketing the proposed mask size while using a feature extraction approach may then be used to
selectthemasksizebasedonthemaskassociatedwiththebestdiscrimination.
111.
X-RAYAPPLICATIONS IN THE FOOD INDUSTRY
X-rays have been usedin the food processing industry for several years. Schatzki
et al. (1 3) demonstrated a technique for determining the density of lettuce heads
al. (14) demonstratedthatx-raytransmission
priortoharvesting.Dieneret
through bruised apple tissue was less than the transmission through nonbruised
tissue, although the typeof bruise was not clearly specified. X-ray machines have
also been used for detecting hollow heart in potatoes (1 5) and split-pit defect in
peaches ( 16). In food processing plants, x-ray-based technologies have success(17). Direct transmission
fully detected stones, bones, metal, and other objects
x-radiology can be useful for evaluating overall internal quality of several food
products. However, the resulting information from these techniquesis unsatisfactory for analyzing specific point locations of a fruit because of inherent volume
averaging effects. The U.S. Department of Agriculture (USDA) now uses linear
array x-ray scanners for detecting food in baggage at some airports.
X-ray technology hasalso been applied for detectionof watercore in apples
(18). Whiledetectionwaspossible,thevariation
of pathlengthorradiation
a very short length at the edge of the fruit to a
through the whole fruit, from
maximum just outside the center, produced an irregularly exposed radiograph.
of
Theapplication of x-radiographyforquantifyingphysicalproperties
fleshy fruits requires appropriate correlations between the physical property
and
x-ray absorption. Tollner et al. (19) present results for apples. Such correlations
enable the useof x-ray CT and lineararray scanners to nondestructively quantify
physical properties of food products. X-ray CT can effectively simulate linear
array scanner action; thus, most discussion is focused on tomography.
a
X-ray CT imaging is a proven method for nondestructively evaluating
cross section of an object. CT greatly reduces volume averaging compared
to
linear array scanner. Each point on a CT image represents a small volume in the
Shahin
146
and
Tollner
scanned plane by the x-ray, while a point on a transmission x-ray image (film)
represents a volume average of many volume elements between the x-ray source
and the film or detector array. X-ray CT enables detailed evaluation of relative
x-ray absorption coefficients over a defined plane within a fruit. One can also
simulate direct transmission (line scan) x-radiography with
a CT scannerby summing results in selected directions. Thus CT scanners are ideal for studying xray absorption properties of fruits.
X-ray CT was used to image the interior region of Red Delicious apples
(19). X-ray absorption properties of
under varying moisture and density states
apples with watercore disorder were different than apples without watercore disto the
order. The increased absorption of x-rays in watercore apples was due
increased water content of the tissue. CT image of watercore apple cross section
clearly revealed the disorder and severity results from the image.
The above studies were
mainly oriented toward manual detection of various
defects from x-ray images. Studies of this type are necessary to determine feasibility of defect detection and to learn the defect signature. X-ray-based defect
detection can at the least provide a machine assist for graders and other quality
control personnel. Once manual defect detection is possible, one may then cona decision
sider automating the feature extraction process and subsequently using
engine to classify the images based on the degree of defect.
IV. FEATUREEXTRACTION
Feature extraction is the applicationof various image processing operations such
as morphological operators, edge detecting with various edge detectors, thresholding, pattern matches, pixel counts, and pixel moments, etc. Operations are not
necessarily in this order, nor do specific protocols include all operations. The
development of the protocol for specific applications is as much an art as it is
science.
One in effect decides at the feature extraction juncture what is and is not
feature noise. The automation process should
begin with considerable input from
those experienced in the physiology of the disorder or defect in question. This
is especially true with issues related to moisture distribution and factors related
to handling and processing, which may affect the distribution of moisture. One
should be familiar with appropriate grade standards and extensionsto these standards, which are typically imposed by individual processors and packinghouse
operators who happen to be the client or customer. One should have experience
in manually extracting the feature of interest before attempting to do same with
computerized feature extraction. Judicious selection
of x-ray energy level and
product orientation relative to the x-ray source is always appropriate. These decisions should be determined in the manual extraction phase.
X-Ray Imaging
147
A. Watercore in Apples
Shahinetal.(20)summarizedacasestudyinvolvingwatercore
in apples.
Watercore is a cosmopolitan disorder that occurs sporadically, prevalent in some
years and absent in others. It occurs only later in the season in mature apples.
In addition, watercore occurs onlyin susceptible cultivars, and generally the characteristic symptoms appear only when the fruit is attached to the tree. According
is described as
to the U.S. grade standard for apples, damage due to watercore
invisible watercore existing around the core and extending to watercore in the
vascular bundles; or surrounding the vascular bundles when the affected areas
surrounding three or more vascular bundles meetor coalesce; or existingin more
than slight degree outside the circular area formed by the vascular bundles; or
any externally visible watercore. A normal fruit has 20-35% of the total tissue
volume occupied by the intercellular air space. In apples with watercore this large
air space is filled with a liquid. The liquid in the air spaces also increases the
density of the fruit. Tollner et al. (19) reported that density of the whole apples
increased with increasing severity of watercore disorder, however, the difference
was not always statistically significant. These changes
in fruit density and absorption characteristics with watercore have been employed in nondestructive detection of disorder. They (19) used x-ray CT to image the interior region of Red
Deliciousapplesundervaryingmoistureanddensitystates.X-rayabsorption
properties of apples with watercore were different from apples without watercore
disorder. The increased absorption of x-rays in watercored apples was due to
increased water contentof the tissue. CT images of watercore apple cross sections
clearly revealed the disorder severity. Results from the image analysis correlated
well with visual observations on the cut-open apples. Throop et al. (21) reported
a classification standard based on natural features for apples to define watercore.
Watercore can be thought of as groups of watery tissues, starting at the
center of the fruit and growing outward. In an x-ray image, the regions corresponding to watercorewill appear as bright regions consisting of adjacent pixels.
This a priori knowledge canbe of great importance to identify features indicative
of watercore. The likelihood principle along
with the laws of perceptual organization can be very useful in determining fruit quality from the image features. This
implies that the area of these brighter regions or blobs might be a good measure of watercore severity. Based on this assumption, fruit area in the processed
image was used as an indicator of watercorein apples (22,23). Since morphological operations are especially good for blob extraction, morphological opening
was used to extract the area features from line scan images
of red delicious apples.
Morphological opening is a combination of erosion and dilation operations. Erosion operation eliminates the false features, whereas dilation operation makes
sure that the true features remain intact. A structuring element approximately the
size of the prefiltering mask is appropriate.
Shahin
148
and
Tollner
Morphological image processing is a powerful tool for examining bloblike shapes and areasin images. McDonald and Chen(24) described some applications where morphological image processing was used for size distribution,
shape discrimination, and texture analysis of agricultural commodities. Throop
et al. (21) presented an algorithm to separate old and new bruised tissues from
stored apples using NIR reflectance. Apple images were first low-pass filtered to
removesmallsurfacecharacteristicsandthennormalized.Theshapes
of the
bruised areas were evaluated based on a shape factor. Morphological closing was
used to eliminate unwanted isolated details. Recognition of agricultural objects
by shape using morphological erosionof the absolute gradient of the x-ray image
was described by Schatzki et al. (25).
The ratio of the fruit areas measured before and after processing may be
an important feature in order to eliminate the effect of fruit size. Variation in the
fruit image intensities may be important due to the fact that watercored fruits are
expected to show more variation in the fruit image as compared to good fruits.
Therefore, standard deviationof the gray values withina fruit image may provide
someinformationaboutwatercoreseverity.Otherapplicationsnodoubtmay
cause other areal features to come to the fore. Typical results before and after
prefiltering plus application of the morphological operator to apple line scan images are shown in Fig. 3.
Image transforms are mainly usedto represent images in compressed form.
This makes the transform coefficients potentially good indicators
of fruit quality.
The Karhunen-Loeve transform (KLT) is the optimum transform that minimizes
the mean square error in the reconstructed image (26). The KLT yields a set of
uncorrelated coefficients, but it is slow to compute. However, the discrete cosine
transform (DCT), which is muchfaster to compute, closely approximates the
information packing ability of the optimal KLT (27). The discrete wavelet transform (DWT) is another commonly used transform in image processing applications. The DWT generates coefficientsthat are local both in frequency and spatial
(28).
domains; a characteristic that the discrete Fourier transform (DFT) lacks
Both DCT and DWT have been widely used in the general areas of image compression and feature extraction applications.
11 was hypothesized that textural features extractable by transforms such
as those above would provide additional information to a decision engine. The
choice of a particular transform i n a given application, however, depends on the
computational resources available and the a1nount of error that can be tolerated.
Information packing ability of DCT is superior to that of DFT and WHT. Although this is usually true for most natural images, the KLT, not the DCT, is the
preferred transform in the information packing sense (29). However, because the
KLT is data dependent, obtaining the KLTbasis images is a nontrivial computational task. For this reason, the KLT is seldom used in practice. Instead, DCT,
whose basis images are fixed, is normally preferred because it is fast and closely
149
X-Ray Imaging
I
Fig. 3 Results of applyingthresholding and morphologicaloperatorto raw line scan
of apples: (a) original line scan image of apples; (b) image following thresholding and
morphological processing.
approximates theKLT. Ahmed et al.(30) first noticed that the KLT basis images
closely resemble the DCT basis images. As the correlation between adjacent pixels approaches unity, the basis images become the same for both the KLT and
DCT (31). The DCT has found widespread applications in speech and image
processing for the purpose of data compression and feature extraction (32) and
Shahin
150
and
Tollner
noise filtering (33). However, no application of DCT for area defect detection
was found in cases of food products.
Wavelets are a set of mathematical functions, called basis functions, that
form a compact description of a signal. These basis functions can be used to
of resolution. As such, wavelets compress
represent signals at multiple levels
images and recover information from even the murkiest images. Wavelets also
isolate the location as well as the scale of features in a signal/image; thus, they
can encode rapidly changing signals in compact form. Details on wavelet transforms are givenby Daubechies (34). Wavelet transform has been used widely for
feature extraction purposes in a variety of applications. Agricultural applications
include grading beef carcasses (35) and picking green peppers in a field of green
leaves (36).
B. Bruises in Apples
Quality defects such as surface bruising decrease market value
of fresh fruits.
of the fresh market fruit producTender fruit varietiesthat make up a large portion
tion are especially susceptible to bruise damage during mechanical harvest and
postharvest handling. According to Studman et al.(37), on the average more than
15% of apples do not qualify as US extra fancy due to bruise damage during
orchard transport. As much as 29% of apples have been reported to have some
bruising after harvest. In order to maintain fruit quality, bruised apples must be
separated during the grading operation. At present, apple sortingis done through
human visual inspection that is inconsistent, inefficient, and subjective. Additionally, bruises in dark-colored fruits such as Red Delicious apples are verydifficult
to see, especially when the fruits are moving on the conveyors, thus further hinderingvisualinspection.Applesarethethirdlargestfruit
crop i n theUnited
States (38). Hence, substantial economic benefits could be achieved by timely
redirection of bruised fruits for alternative processing, proper utilization of storagespaceforstoringonlygoodfruits,andreducinglaborcostsformanual
sorting.
Impact stress occurs when the productis dropped from a sufficient distance
to cause injury. This type of injury is generally seen as bruising. With bruising
the injury is restricted to the interior flesh of the tissue and in many products
may be only initially detected as a water-soaked area after peeling.
With time,
in
the exposure of the damaged cells to air in the intercellular spaces results
thetypicalbrowning
symptoms. The damaged tissue mayeventuallybecome
desiccated, thus opening the possibility
of x-ray imaging. X-ray imaging
is a
leading contender becauseof its abilityto detect both external and internal defects
such as surface bruises and watercore (39,40). Apple defect classification from
x-ray images warrants major emphasis. Shahin et ai. (41) found that edge detection in the zone just under the fruit skin correlated with bruise presence and
X-Ray imaging
151
severity. Figure 4 shows line scan images of apples prior to prefiltering (Fig. 4a)
and after prefiltering and Roberts Edge operators were applied (Fig. 4b). Figure
4c shows the effect of substituting the median filter for the Gaussian filter. The
Gaussian filter was ultimately chosen based on subsequent testing. The edge corresponding to the skin edge was removedby further image subtraction operations
(results not shown).
C.Diseases
in Onions
Sweet onions are an important specialty crop grown in the Vidalia region of
of onions,sweetonionslackpungencycomGeorgia.Unlikeothervarieties
pounds (42). As a result, they are readily susceptible to disease inoculation and
tend to have a shorter shelf life. Several decay-causing organisms can possibly
(43,44). Mechanical injuries
invade the onion bulbs at any time via soil or air
occurring during spraying and harvest operations may damage the bulbs leading
to infection. Quality defects such as internal decay, ring separation, and internal
of onions. These physiological changes cause
sprouting decrease the market value
voids and/or other tissue change, which alters moisture distribution. X-ray imaging has shown promising results for detecting internal defects
in fruit and vege(45), pests (39), watercore
tables such as interior voids and foreign inclusions
(20), and freeze damage (46).
X-ray imaging and image processing techniques were employed (47) to
identify poor quality onions. They emphasized theneed for better image analysis
techniques for improved accuracy of classification. In light of the physiology of
the onion bulb and physics of the line scan operation, edge detection techniques
are expected to perform well for enhancing internal defect features in x-ray imin eggs (48) and bruise
ages. Success of edge detectors for enhancing cracks
damage in apples (49) provided a reasonable likelihood of success with disease
detection in onions (50) sincethe feature of interest in each case appearsin images
as a line or edge. Figure 5 shows the results of line scanning onions before (Fig.
Sa)andafter(Fig.5b)prefilteringandfeatureextraction.Arealfeaturesplus
various transform coefficients provide a suite
of potential inputs to a decision
engine for classification purposes.
D.MiscellaneousStudies
Long-established feature extraction approaches for optical images (e.g., see Ref.
51) may be applied to x-ray images. X-ray image feature extraction research
is
of baby
ongoing at several laboratories. Problems including the sex separation
chicks (52), insects, bruises, rots and browningof apples (39), and other general
agricultural products ( 5 3 ) are being researched. Techniques ranging from simple
segmentation to sophisticated watershed delineation algorithms (54) are under
152
Tollner and Shahin
Fig. 4 Line scanimage of apples (a) followed by application of Gaussianfilterplus

Roberts edge detection (b) and application of median filter plus Roberts edge detection
(c).The Gaussian filter proved superior to the median filter in later tests.
X-Ray Imaging
153
Fig. 5 Line scan image of onions before and after Gaussian filtering and Roberts edge
detection. (a) Original line scan image of onions: dark features indicate defects. (b) Processed image showing features of good and defective onions:bright arcdlines insidebulb
boundary indicate defects.
Tollner and Shahin
154
evaluation. The numberof laboratories for x-ray imaging research for food products continues to expand.
V.
DECISIONENGINES FORFEATURECLASSIFICATION
Pattern classificationis the taskof recognizing the class label

w from the measurement vector v. This task is equivalent to establishing a many-to-one mapping x
+ d from the measurement space V into the decision space D, containing M
discrete points, each of which represents one of the M classes. A common task
in pattern classification is variable or feature selection. Typically,
a single feature
has little discriminating power. Hence, combining several features improves the
chances of coming to a satisfactory result. But combining many features makes
it more likely that the feature set as a whole has correlation and redundancy. In
this situation, it is not very useful to evaluate the potential contributionof a single
feature in isolation to solve the classification task. Rather, mutual dependencies
among the features need to be considered. The decisive task is to determine the
all features is already
contribution of a single feature after a certain subset of
used for classification. The answerto this problem is rank ordering of the features
based on their additional contribution after a subset has already been selected
(55). The quality of the many-to-one mapping among the features truly related
to the defect affects the classification noise. For example
a nonlinear many-to-one
a linear mapping.
mapping when needed would have less classification noise than
Statistical methods, suchas Bayesian classifiers, can provide optimal classification, but their performance deteriorates when the input data fail to meet the
assumptions of normality and equality of covariance matrices. Artificial intelligence approaches such as fuzzy logic and neural classifiers, on the other hand,
entail less stringent assumptions about the statistical characteristics of the input
to perform
data. Thus, fuzzy logic and neural network classifiers are expected
equally good or better than the Bayesian classifier.
A.
Bayesian Classifier
The Bayesian classifier is a probabilistic approach

to recognition. Probability
in patternrecognitionbecause
of theranconsiderationsbecomeimportant
clasdomness under which pattern classes normally are generated. The Bayesian
sifier is optimal in the sense that, on average, it yields the lowest probability of
committing classification error. The Bayesian classifier is difficult to apply in
practice, especially if the number of representative patterns from each class
is
not large or if the underlying form of the probability density function is not well
behaved. By far the most prevalent form assumed
is the Gaussian probability
X-Ray Imaging
155
density function. The closer this assumptionis to reality, the closer the Bayesian
classifier approaches the minimum average loss in classification. The Bayesian
classifier implements a decision function of the form:
dj(x) = p(x/oj)P(oj)
1, 2, . . . , M
where a pattern x is assigned to class mi if di(x) > dj(x) for all j f i. P(wj) and
p(x/oj) are the probability of occurrence and the probability density function of
class w,, respectively. M is the number of classes the pattern can be assigned to.
Assuming a Gaussian probability density function, the Bayes decision function
is given by:
d,(x) = In P(wj) - 0.5 lnlCjl - 0.5[(x - rn,)TC;(x
mj)]
(4)
where mj and C, are the mean vector and covariance matrix, respectively.
In
addition to the normality assumption, the Bayes discriminant function assumes
that the group covariance matrices are equal and the output changes linearly with
the input variables.
Outputs from typical discriminant analyses enable one to analyze the relative contribution of the various inputs. Inputs not contributing meaningful information need not be further considered, thus simplifying the model. Datasets for
discriminant analyses are comprisedof the image features plusan expert classification using destructive techniques. Model development generally is performed
using a subset of the data.The remainder of the data are used for model validation.
The primary outputof interest is the classification of each sample product. Percent
accuracy of the classification along with the percent of false positives are generally reported. Classification accuracy may range anywhere from random (e.g.,
number of categories divided by 100%;very poor) to 90% plus, considered to be
very good. In cases where the expert standardis hand inspection, 90% accuracy is
to assure
considered excellent. Percent accuracy represents the systems ability
a good product, and false positives (100% accuracy) represent a tangible loss to
the system in that a good product is rejected.
The significant image features found to be significant contributors
in an
apple watercore study were blob area, mean grey value, and the tenth discrete
depending
cosine coefficient (20). Prefiltering may or may not be appropriate,
on the feature of interest. Selection of the prefiltering mask can be completed
once one has determined the significant image features contributing
to meaningful
classification. In the case of apples with watercore, Shahin et al. (20) justified
the size of the structuring element used in the morphological processing based
on the image noise statistics (20). Results from the watercore study are summarized in Table 1 . For the most part the classification was generally acceptable by
modem standards. Results for the moderate watercore category were worse
than either the none or the severe category. Some of the moderate cate-
156
Tollner and Shahln
Table 1 Classification Results for Bayes Classifier Applied to Red Delicious Apples
with Watercore
No/mild
Moderate
Severe
Total predicted
False
positives
Predicted class
actual
Watercore
No/mildclass
Moderate
Severe
10
4
0
52
m;
2
19
19
actual
(%)
(%I
58
14
63
1618
90
83
50
89
21
Source: Ref. 49.
gory error was thought to

moderate category.
be related to the difficulty of hand classifying the
6. Fuzzy Logic
Fuzzy logic is based on fuzzy set theory introduced by Zadeh
(56). Fuzzy logic
is a nonparametric decision engine that can encompass nonlinear relationships.
Biological systems are complex in nature, and when idealized to develop representation by traditional models, they can distort to the extent where the models
are not representative and have little
use. Fuzzy logic methods present alternatives
for modeling complex biological systems
with a remarkably simpleway of drawing precise conclusions from vague, ambiguous, or imprecise information.
The initial stage of fuzzy model development depends to some extent on
intuition and the designers feel for the nature of the problem. Design
of a
fuzzy model does not require a well-defined mathematical model of the system
to be modeled. It can start with a general idea of how the model should work.
The idea may consistof defining input variables and their membership functions
(fuzzification), a set of rules to define the model behavior under different conditions (fuzzy inferencing unit,FIU), and an output variable to translate fuzzy output of the FIU into a crisp value (defuzzification).
Several fuzzy models have been developed to predict the response of biological systems. Among several applications, fuzzy logic has been used for predicting yield for prescription farming (57), estimating forest parameters from
image data (58), analyzing plant structure(59), diagnosing tomato diseases(61),
predicting peanut maturity (60), and sorting tomatoes(62) and apples (23) based
on quality.
The dataset for fuzzy classifier development consists of image features plus
expert data. Fuzzy approaches do not enable a ready evaluation of significance
X-Ray
157
of contribution by each parameter. Thus we suggest using Bayesian analyses on

the model building dataset for purposes of determining the relevant parameters
fortheFuzzyclassifierdevelopment.Fuzzyclassifierdevelopmentwhenthe
number of inputs exceeds twois hindered by difficulties in tuning the membership
functions for optimal performance. Shahinet al. (20) developed a fuzzy classifier
for apple watercore using the three significant image features (area, mean grey
value, tenth discrete cosine coefficient) identifiedin the Bayesian analyses above
and found results very comparable to those presented in Table I .
C.NeuralNetworks
Artificial neural networks (ANNs) provide an alternative way of modeling sysor ill-definedinput-outputrelationships.ANNclassifiers
temswithunknown
in a manner similar to humans. The
have the capability of classifying patterns
power of a neural network lies in its ability to model nonlinear relationships and
in realits robustness to deal with noisy and incomplete data commonly found
life situations. ANNs are capable
of forming highly nonlinear decision boundaries
and have an intrinsic power to generalize. Neural networks look for patterns in
develop the ability to correctly
training sets of data,learnthesepatterns,and
classify new patterns.
The multiple-layer feedforward network (MLFN) was originally described
byRumelhart et al.(63). It is themostwidelyusednetworkarchitecturefor
problems that can make use of supervised training (learning through examples).
The MLFN can require longer training periods, but execution of the trained network takes relatively little time. Hence, real-time applications are best served by
the MLFN model (64). Details of neural networks can be found in a number of
publications (65-67).
Neural networks have been successfully used for food quality evaluation
and related applications. Boudolf (68) used the neural network approach to predict peanut maturity from nuclear magnetic resonance (NMR) data.
Pate1 et al.
(48) successfully developed a neural network for crack detection in eggs using
computer vision. Elizondo et al. (69) developed
neural
a network model to predict
flowering and physiological maturity dates of soybean. Das and Evans (70) used
computer vision and neural networks to separate fertile and infertile eggs. Deck
et al. (71) compared neural networks and statistical methods for sorting potatoes
into various classes using machine vision.
The dataset for theneural net classifier development is identical to that for
the Bayesian analyses except that insteadof a model building set and a validation
data set, the data are divided into model building, model testing, and validation
datasets. The MLFN requires that a test set be accessed for limited testing in the
model building phase. Using the same three image features used with the Bayesian and fuzzy classifier development discussed above, Shahin et al. (20) found
Tollner and Shahln
158
Table 2 Classification Results for Neural Classifier with Four Hidden Nodes
Predicted class
Watercore observed
No/mild
class
Nolmild
Moderate
Severe
Total predicted
False
Total
Accuracy positives
observed
(%)
(%I
Moderate
-3
3
0
58
Severe
- 3
14
2
13
58
95
57
5
38
18
19
90
Observed watercore seventy was assigned based on visual judgement of the cut fruit.
substantial improvement in classifier accuracy and false-positive scores (Table

2), especially in the moderate category. Similar studies with old bruises
in apples
(41) and diseases in onions (50) resulted in improved classification accuracy.
D. OtherClassifiers
Bayesian, fuzzy, and neural net classifiers tend to represent classifier types
in and
themselves do not comprise the entire universe
of classifiers. A feature extraction
methodcalledthemaximumrepresentationanddiscriminatingfeature
(MRDF)was reported(72) to classify good and bad pistachio with
nuts accuracies
be an advanced nonlinear implementaat or above90%. Their method appears to
tion of the Bayesian method. They found that adding the quadratic component
to the linear model increased accuracy, which may explain why a number
of
are
studies (40,41,50) reported increased accuracy with neural nets. Neural nets
known to represent nonlinear data fairly well.
VI. SUGGESTED PROTOCOL FOR CLASSIFIER

DEVELOPMENT
Development of a protocol for obtaining usable
x-ray images and optimal extraction of usable features for defect identification has been the focus of this chapter.
The following steps appear to provide a protocol for developing classifiers from
x-ray images:
1. Select the x-ray energy level(s) depending on product size, density,

and available equipment.
2. Run preliminary studies for optimizing orientation of product. Determine the characteristics of the features that correlate to defects (e.g.,
X-Ray Imaging
159
blobs, lines, etc.). Determine the nature of features that

do not correlate
to defect. Select an image processing method that minimizes undesirable and enhances desirable features.
Areal features,transformfeatures, and other common image processing features should be considered.Considerrevisitingequipmentselection
if multipleenergies,
image intensifiers, etc. would result in enhanced features.
3. Select some test materials, such as acrylic, and determine the structure
of the correlation functions within the image. Determine if artifact removal by preprocessing would be beneficial. Determine
if morphological processing is appropriate based on persistence of correlation. Size
of mask may be estimated based on the noise present and correlation
persistence.
100 or more typically) good and defect product at
4. Scan multiple (e.g.,
theselectedorientationandenergylevels.Runtheselectedfeature
extraction processes.
5. Develop a Bayesian classifier for initial parameter evaluation, using a
subset of the data for model building. Using the significant factors,
of
evaluate a range of filter masks that bracket predicted mask sizes
step 3.
6. Train a neuralnetclassifier (or othernonlinearclassifier if desired)
using the selected inputs from step 5 and evaluate using the validation
data set. Evaluate the results in the light of expected accuracy, quality
of expert data, and client expectation. To improve resultsmay
one have
to revisit step 1.
VII.
FUTURETRENDS
X-ray technology has been usedin the past in the food industry for foreign matter
detection,such as screeningmeat/poultryproductsforextraneousmaterials.
However, single energy x-ray technologyis incapable of identifying small defects
and contaminants. Small defects and contaminants are of considerable interest
(73) because of lack of availability of sufficient number of qualified inspectors.
This is often because the variance in mass density between the defect/contaminant and the productand the product noise confound positive identification. Adding an additional scanning energy is being proposed to address these concerns.
Dual energy scanning allows imaging based on molecular weight and mass
density. One could use images at each energy level to further enhance features
of interest. Simultaneous measurementof moisture and soil density has been done
using dual energy level monochromatic sealed gamma sources for many years.
Aylmore (74) developed a dual energy gamma CT system, which successfully
detected plant moisture uptake and root growth. They developed an image or
160
Tollner and Shahin
map of total absorption coefficients at each energy level.For each corresponding

image pixel, they wrote Eq. (2) for each of two energy levels, knowing the dry
soil and water absorption coefficients at each energy level. They simultaneously
solved the equations for the volumetric moisture content and dry bulk density.
The x-ray sources are polychromatic. This requires detectors that can detect
x-rays at multiple wavelengths (i.e., energy levels, 140 kV vs. 70 kV) to enable
effective results. Dual energy x-ray technology has been developed
by Ergun and
in medicalanalysis(76-79).The
Mistretta(75).Ithasbeenusedeffectively
major medical application is bone densitometry.
Line scanners used for security purposes now use multiple x-ray energy
level-based images for highlighting features of interest. They typically use one
polychromatic source with detector arrays with 70 and 140 kV energy windows.
in processedfoodsand
Use ofthistechnologyfordetectingforeignobjects
diseasesldefects in fresh foods is just beginning.
One issue not researched as yet is that of the required residence time of
product in the detector. The residence time would be the total time required for
irradiation plus time required for image processing, feature extraction, classification decision, and action actuation. This cannot as yet be projected until working
prototypes incorporating hardware processing for each step in the classification
protocol are available for testing.
The information gathered from this study could be useful in developing a
decision support system such as the one shown in Fig. 1. The scheme shown in
NIR sensors to
Fig. 1 may be conceptually expanded to include optical and/or
give external quality attributes. Needs of the client and general economic conditions must be fxtored into a more complete classifier. Current trends in labor
availability and economics will likely continue to drive the industry towards more
automated inspection and decision support systems.
REFERENCES
JA Richards, FW Sears, MR Wehr,MWZemansky.ModernUniversityPhysics.
Reading, MA: Addison-Wesley, 1960.
2. K Morita, Y Ogawa, CN Thai. Softx-ray imaging for detection of foreign materials
in foods. Paper No. 986067.St. Joseph, MI: American Society of Agricultural Engineers (ASAE), 1998.
3 . M Lefebvre. Potato Operation: Automatic Detection of Potato Diseases. SPIE Proc.
Optics in Agriculture, Forestry and Biological Processing. Boston: SPIE
1994, pp
I.
2-8.
4. NK Gupta. X-ray system for on-line detection of foreign objects in food. Presented
at the Food Processing Automation Conference IV, Chicago, IL, November 1995.
5. TF Schatzki. Personal communication, 1998.
X-Ray
161
6. RE Garrett, DH Lenker. Selecting and sensingx and gamma rays. In: Quality Detection of Foods. St. Joseph, MI: American Society of Agricultural Engineers, 1976,
pp 107-115.
7. EB Goldstein. Sensation and Perception. 4th ed. Pacific Grove, CA: Brooks/Cole
Publishing Co., 1976.
8. RW Hamming. Digital Filters. Englewood Cliffs, NJ: Prentice Hall, 1977.
9. AC Kak, M Slaney. Computerized Tomographic Imaging. New York: The Institute
of Electrical and Electronics Engineering, Inc., 1988.
IO. IC Russ. The Image Processing Handbook, 2nd ed. Ann Arbor: CRC Press, Inc.,
1995.
11. FL Barnes, SG Azavedo, HE Martz, GP Roberson, DJ Schneberk, MF Skeate. Geometric effects in tomographic reconstruction. Lawrence Livermore National Laboratory Rep. UCRL-ID- 10.5 130, 1990.
12. MA Shahin, EW Tollner, SE Prussia. Filter design for optimal feature extraction
from X-ray images. Trans ASAE, 1998.
13. TF Schatzki, SC Witt, DE Wilkins,DH Lenker. Characterization of growing lettuce
from density contours: I . Head Selection. Patterns Recog 13(5):333-340, 1981.
to sort bruised
14. RG Diener, JP Mitchell, ML Rhoten. Using an X-ray image scan
apples. Agric Eng 51(6):356-357, 361, 1971.
15. RE Nylund, JM Lutz. Separation of hollow heart potato tubers by means of size
grading, specific gravity and X-ray examination. Am Potato J 27-214-222, 1950.
16. SV Bowers, RB Dodd, YJ Han. Nondestructive testing to determine internal quality
of fruit. ASAE Technical Paper No. 88-6369. St. Joseph, MI: American Society of
AgriculturalEngineers,1988.
17. GL Gerber, QA Holmes, R Calhan. Industrial machine vision with X-ray sensor for
online food processing inspection. SME Paper MS85-1009. Dearborn, MI: Society
of Manufacturing Engineers, 1985.
18. BL Upchurch, HA Affeldt, WR Hruschka, JA Throop. Optical detection of bruise
and early frost damage on apples. ASAE Technical Paper No. 89-3013. St. Joseph,
MI: American Society of Agricultural Engineers, 1989.
SE Prussia.RelatingX-rayabsorptionto
19. EW Tollner, YCHung,BLUpchurch,
density and water content in apples. Trans ASAE 35(6): 1921-1928, 1992.
20. MA Shahin, EW Tollner, HR Arabnia and MD Evans. Watercore featuresfor sorting
red delicious apples: a statistical approach. Trans ASAE, 1998.
21. JAThroop, DJ Aneshansley,BLUpchurch.DetectingWatercoreinAppleswith
Machine Vision. ASAE Paper No. 91-7536. St Joseph, MI: ASAE, 1991.
22. MA Shahin, EW Tollner. Detection of watercore in apples using X-ray linescans:
Feature extraction and classification. Proc NRAES Conference on Sensors for NondestructivE Testing, Orlando, FL, February 18-21, 1997, pp 389-400.
23. MA Shahin, EW Tollner. Apple Classification based on watercore features using
fuzzy logic. ASAE Paper No. 97-3077, St. Joseph, MI: ASAE, 1997.
24 T McDonald, Y Chen. Applicationof morphological image processingin agriculture
Trans ASAE 33(4): 1345- 1352, 1990.
25. TF Schatzki, A Grossman, R Young. Recognitionof Agricultural Objects by Shape.
IEEE Trans Pattern Anal Machine Intell 5(6):645-653, 1983.
26. WK Pratt. Digital Image Processing. New York: John Wiley and Sons, 1978.
162
Tollner and Shahin
27. WKou.DigitalImageCompressionAlgorithmsandStandards.Norwell,MA:
Kluwer Academic Publishers, 1995.
28. N Sweeney. Wavelet transforms represent signals in terms of both time and scale.
Personal Eng 37-42, 1996.
29. HP Kramer, MV Mathews. A linear coding for transmitting a set of correlated signals. IRE Trans Inform Theory T-2:41-46, 1956.
30. N Ahmed, T Natarajan, KR Rao. Discrete Cosine Transforms. IEEE Trans Comp
C-23~90-93, 1974.
31. RJ Clark. Transform Coding of Images. New York: Academic Press, 1985.
32. M Bi, SH Ong, YH Ang. Coefficient Grouping Method for Shape-Adaptive DCT.
Electron Lett 32(3):201-202, 1996.
33. AH Lia, NH Yung. New feature-preserving filter algorithm based on a priori knowledge of pixel types. Optical Eng 35( 12):3508-3521, 1996.
34. I Daubechies. Wavelet bases. In: L Schumaker, ed. Recent Advances
inWavelet
Analysis. Boston: Academic Press, 1994, pp 237-257.
35 ND Kim, V Amin, D Wilson, G Rouse,S Upda. Texture analysis using multiresolutionanalysisforultrasoundtissuecharacterization.
Rev Prog QuantNondestruct
Eva1 2: 1351 - 1358, 1997.
36. C Davidson, D Rock. Wavelets and HONN: plx-perfect marriage. AI Expert (January): 31-35,1995.
37. CJStudman,GKBrown,
EJTimm,NLSchultz,MJVreede.Bruisingonblush
and
non-blush
sides
in apple-to-apple
impacts.
Trans
ASAE
40(6):
16551663,1997.
38. RSChilds,RAMilligan,GBGWhite,WC
Stiles. Adynamicapproachtoapple
orchard replacement. Cornell University Agricultural Economics Research
83-1 I .
Ithaca,NY:CornellUniversity, 1983.
39. TF Schatzki, RP Haff, R Young, 1 Can, LC Lee, N Toyofuku. Defect detection i n
apples by means of X-ray imaging. Trans ASAE 40(5): 1407- 141 5, 1997.
40. MA Shahin, EW Tollner, SE Prussia. Detecting defects in apples using X-ray imaging: Filter design and features enhancement considerations. TransSt. Joseph, MI:
ASAE,1998.
41 MA Shahin, EW Tollner, RW McClendon, HR Arabnia. Apple classification based
on surface bruises using image processing and neural networks. Manuscript submittedtoAI Applications, 1998.
42. BW Maw, YC Hung, EW Tollner. Some physical properties of sweet onions. ASAE
Tech. paper no. 89-6007. St. Joseph, MI: ASAE, 1989.
43. DR Sumner. Soilborne pathogenic fungi, root diseases and bulb rots. In: BW Maw,
ed. Georgia Onion Research Extension Report. Cooperative Extension Service, University of Georgia, College of Agriculture and Environmental Sciences, 1994.
44. RD Gitaitis. Bacteriology results 1993 onion trials. In:BW Maw, ed. Georgia Onion
Research Extension Report. Cooperative Extension Service, University of Georgia,
College of Agriculture and Environmental Sciences, 1994.
45. EW Tollner, HA Affeldt Jr, GK Brown, P Chen, N Galili, CG Haugh, A Notea, Y
Sarig, T Schatzki,I Shmulevich, B Zion. Nondestructive detectionof interior voids,
foreign inclusions and pests. Nondestructive technologies for quality evaluation of
fruits andvegetables,Proceedings
of theInternationalworkshopfunded
by the
X-Ray Imaging
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60
61
62.
63.
163
United States-Israel binational agricultural research and development fund (BARD),

Spokane, Washington. St. Joseph, MI: ASAE, 1993.
BL Upchurch, HA Affeldt, DJ Aneshansley, GS Birth, RP Cavalieri, P Chen, WM
Miller, Y Sarig, Z Schmilovitch, JA Throop, EW Tollner. Detection of internal disorders. Nondestructive technologies for quality evaluation of fruits and vegetables.
Proceedings of the International Workshop Funded
by the United States-Israel Binational Agricultural Research and Development Fund (BARD), Spokane, Washington. St. Joseph, MI: ASAE, 1993.
EW Tollner, YC Hung, B Maw, DR Sumner, RD Gitaitis. Nondestructive testing
for identifying poor quality onions. Proc SPIE Tech. Conf. 2345:392-402, 1995.
VC Patel, RW McClendon, JW Goodrum. Crack detection in eggs using computer
vision and neural networks. AI Appl 8(2):21-31, 1994.
MA Shahin, EW Tollner, RW McClendon. AI classifiers for sorting apples based
on watercore. Manuscript prepared for AI Applications, 1998.
MA Shahin, EW Tollner, RD Gitaitis, DRSummer,BWMaw.Classification
of
onions based on internal defects using image processing and neural network techniques. Trans ASAE, 1999.
AK Jain. Fundamentals of Digital Image Processing. Englewood Cliffs. NJ: Prentice-Hall,1989.
Y Tao. Feature extraction and pattern recognition method for automated sex separation of baby chicks. Tech. Paper No. 983046. St. Joseph, MI: ASAE, 1998.
DA Casasent, A Talukder, HW Lee. X-ray agricultural product inspection: Segmentation and classification. Proc SPIE 3205:46-55, 1997.
A Talukder, D Casasent,HW Lee, PM Keagy, TF Schatzki. Modified binary watershed algorithm for segmentation of X-rayed agricultural products. Proc SPIE 3543:
1998.
C McHenry. Multivariate Subset Selection. J Royal Stat Soc C 27(23):29 1-296,
1978.
LA Zadeh. Fuzzy sets. Inform Control 8338-353. 1965.
JR Ambuel, TS Colvin, DL Karlen. A fuzzy logic yield simulator for prescription
farming. Transactions of the ASAE, Vo137(6): 1999-2009, 1994.
FC Maselli, T de Filippis Conese,
S Norcini. Estimationof forest parameters through
fuzzy classification of TM data. IEEE TRANS. Geoscience Remote Sensing, Vol
33(1):77-83, 1995.
W Simonton. Bayesian and fuzzy logic classification
for plant structure analysis.
ASAE Paper No. 93-3603, ASAE, 2950 Niles Rd, St Joseph, MI 49085-9659, 1993.
MA Shahin. Fuzzy logic model for predicting peanut maturity. Submitted to Transactions of the ASAE, 1998.
C-H Wu, S-W Hu. Fuzzy related computer vision for diagnosing tomato diseases.
ASAE Paper No. 933033, ASAE, 2950 Niles Rd,St. Joseph, MI 49085-9659, 1993.
N Pu. Fuzzy logic DSS for sorting tomatoes based on quality. Unpublished Masters
Thesis, The University of Georgia, Athens, 1994.
DE Rumelhart, GE Hinton, JL McClelland, eds. A general framework for parallel
distributedprocessing.In:ParallelDistributedProcessing-Explorationsinthe
Microstructure of Cognition.Vol. 1: Foundations.Cambridge,MA:MITPress,
1987.
Shahin
164
and
Tollner
64. TMaster.SignalandImageProcessingwithNeuralNetworks.NewYork:John
Wiley and Sons, Inc., 1994.
65. JL McClelland, DE Rumelhart. Eds. Parallel Distributed Processing: explorations
in the microstructure of cognition. Vol. I Foundations. Cambridge, MA: MIT Press,
1986.
66. R Hecht-Nielsen. Neurocomputing. New York: Addison-Wesley, 1990.
67. DF Cook, ML Wolfe. A back-propagation neural network to predict average air
temperature. AI Appl 5( 1):40-46, 1991.
68. VA Boudolf 111. Predicting peanut maturity: an artificial neural network approach.
Unpublished masters thesis. Athens: The University of Georgia, 1996.
69. DA Elizondo, RW McClendon, G Hoogenboom. Neural network models for predicting flowering and physiological maturity of soybean. Trans ASAE 37(3):98 1 988,1994.
70. KC Das, MD Evans. Detecting fertility of hatching eggs using machine vision
11:
Neural network classifiers. Trans ASAE 35(6):2035-2041, 1992.
71. S Deck,CTMorrow. PHHeinemann,andHJSommer
111. Neuralnetworksfor
automated inspection of produce. ASAE Paper No. 92-3594.St. Joseph, MI: ASAE,
1992.
72. A Talukder, D Casasent, H-W Lee, PM Keagy, TF Schatzki. A new feature extraction method for classification of agricultural products from X-ray images. SPIE Proceedings for Photonics East, Boston, 1988.
73. USDA. Food Safety Research: Current Activities and Future Needs, Washington,
DC, 1994, p 36.
74. LAG Aylmore. Use of computer-assisted tomography in studying water-movement
around plant-roots. Adv Agron 49: 1-54, 1993.
75. DL Ergun, CA Mistretta. Single exposure dual energy computed radiography: improved detection and processing. Radiology 174:243-249, 1990.
76. MJ Bentum, RGJ Arendsen, CH Slump, CA Mistretta, WW Peppler and FE Zink.
Design and realization of high speed single exposure dual energy image processing.
Proceedings of the Fifth Annual IEEE Symposium on Computer-Based Medical SYStems Durham, North Carolina, 1992, pp 25-34.
77. DL Ergun,WW Peppler, JT Dobbins111, FE Zink, DG Kruger, F Kelcz,
FJ de Bruijn,
EB Nellers, Y Wang, RJ Althof, MGJ Wind. Dual-energy computed radiography:
improvements in processing. Medical imaging: image processing. Proc SPIE 2167:
663-67 1, 1994.
DG Johnston. Compari78. SA Beshyah, C Freemantle, E Thomas, B Page, M Murphy,
son of measurements of body composition by total body potassium, bioimpedance
analysis, and dual-energy x-ray absorptiometry in hypopituitary adults before and
during growth hormone treatment. Am J Clin Nutr 61:1186-1194, 1995.
79. SA Jebb, GR Goldberg, G Jennings, M Ella. Dual-energy x-ray absorptiometry measurements of body composition: effects of depth and tissue thickness, including comparisons with direct analysis. Clin Sci 88:319-324, 1995.
5
Nuclear Magnetic Resonance
Techniques and Their Application
in Food Quality Analysis
R. Roger Ruan and Paul
L. Chen
University of Minnesota, St. Paul, Minnesota
1.
INTRODUCTION
Laboratoryanalysis
of foodqualityofteninvolvesextensivepreparation
procedures,duringwhichfoodsareseverelydisturbed
by size-reducing,deforming, or diluting steps. This
is very unlike the way
in which food consumersevaluatethequality
of foodproducts (Le., foodsareconsumedwhile
theirintegrity is generallyintact.).Onemaythereforeask:Arethecurrent
chemical and physical methods for food quality measurement reliable? Naturin a completely
ally it would be desirable to be able to analyze food quality
nondestructiveandnoninvasiveway.Nuclearmagneticresonance(NMR)
a task.NMR
techniquesareamongthosemostcapableofperformingsuch
techniques,includingNMRspectroscopyandNMRimaging,ormorefreto nondequentlytermedmagneticresonanceimaging(MRI),canbeused
structivelyandnoninvasivelystudythechemicalandphysicalpropertiesand
phenomena, anatomical structure, and dynamic processes of food materials and
products.
In this chapter we will introduce the basic principles and methodologies
of proton NMR spectroscopy and imaging techniques and present some examples
of practical applications of the techniques.
165
166
II.
Ruan and Chen
BASIC PRINCIPLES OF NMR SPECTROSCOPY
A. NuclearMagneticResonance
N M R is a spectroscopic technique based on the principleof resonance. Certain

nuclei (spin quantum number I # 0) are considered as spinning about an axis
(Fig. 1). The spinning generates a small magnetic field, known as magnetic moment. Like a normal magnet bar, this magnetic moment has a north and a south
pole, called a nuclear magnetic dipole. When the nucleus is placed in a static
magnetic field, itwill interact with the field via its dipole, that is, the dipole will
tend to align with the field much as a compass needle aligns with the earths
magnetic field. In addition to this magnetic dipole, a second phenomenon occurs
when the nucleus is placed in the static magnetic
field. This canbe demonstrated
by considering the spinning nucleus as a spinning top. When the top is spun, it
of the earths
will spin, but it begins to wobble when it slows down because
gravitational field. This wobbling or angular spinning is called precession. Similarly, the spinning nucleus in a static magneticfield will precess about the static
of the nuclear dipole is called
Larmor
magnetic field. The frequency of precession
frequency (1) defined by the M o r equation:
o = ?Bo
(1)
where o is the frequency of precession,Bo is the strength of the static magnetic

field, andy is termed the geomagnetic ratio and has a precise value characteristic
of each nuclear species. For example, a proton has a y of 42.6 MHz/T. If the
static field Bo is 1 T, the o is 42.6 MHz. If Bo = 4.7 T, 61is -200 MHz.
However, the orientationsof magnetic dipolesin a static magnetic field are
different depending on their magnetic quantum number, that is, some nuclei align
Fig. 1 Magnetic moment (p)precesses about the bulk magnetic field Bo with an angle 8.
NMR in Food Quality Analysis
167
themselves in either the same direction or the opposite direction as the field. In
the classic view described above, the alignment of the dipoles is determined by
whether the nuclei precession is in a clockwise or counter-clockwise direction.
A nucleus that has its spin aligned with the
field will have a lower energy and
more stable than when it has its spin aligned in the direction opposite to the field
(Fig. 2 ) .
The energy difference BE is proportional to the strength of the static field
and the population difference:
and
and N,,,,, represent the

where R is the Planck's constant divided by 27c, Nupper
population of nuclei in the upper and lower energy levels, respectively, k is the
Boltzmann constant, and T is absolute temperature. Equation (3) is also called
Boltzmann's law. which describes the thermal equilibrium of the nuclear spins.
One NMR experimentinvolvesinducingtransitionbetweentheneighboring energy levels by absorption or emission of a photon with the requisite
energy. This requisite energy is applied in the form of a rotating magnetic field
B , or radiofrequency (RF) pulse, which exactly matches the Larmor frequency
and has an energy equalto the AE, causing resonant absorption or emission of the
<
-112
........................
Aligned anti-parallel
Zen, field
2:w
ti
........................
Aligned parallel with B,
Fig. 2 NuclearZeemansplitting
quantum number I/:.
of energy levels in a magneticfield
for spins with
Ruan and Chen
168
energy by the nuclei. This resonance effect is therefore termed nuclear magnetic
resonance.
B. VectorDiagrams of Magnetization
The magnitude of the magnetization caused by the RF pulse can be visualized
through the vector model (Fig. 3). For protons with I = %I/?, the magnetic quantum number m I takes two values, + I / ? and --I/?, that is, the spins align in either
the same or opposite direction with Bo. Thenumber of spins pointing in the same
direction as Bo is greater than that pointingin the opposite directionas B,)because
the former alignment represents a low-energy or stable state. The individual spins
precess about the Bo at the Larmor frequency. The net magnetization of the spin
system is M. The length of the vector M corresponds to the difference in population between the two energy states.
We can put this vector model in a laboratory coordinate with a set of axes
as shown in Fig. 4. The applied magnetic field Bo is traditionally parallel to the
z-axis, and the net magnetization, M,, also points in the z-axis direction. If the
spin system remains in the Bo field long enough, the number of spins oriented
with the field reaches an equilibrium value. This value is termed Mo and is the
largest possible value of magnetization.
Because the net magnetization is precessing so fast, visualizing its motion
is difficult. One solution to this problem is to use a rotating frame to replace the
stationary frame. The condition is that the rotating frame rotates about the z-axis
at exactly the Larmor frequency. Therefore, within the rotating frame, the net
Fig. 3 Vector modelof magnetization of protons in the presenceof a permanent field Bo.
169
Fig. 4 Magnetizationvectorsin a stationarylaboratorycoordinatesystem.
magnetization does not appear to be precessing and is thus easier to follow. The
rotating frame uses x, y, and z to label the three axes as shown in Fig. 5.
Visualization of interactions between applied RF field B , (in a direction
perpendicular to Bll)and spin magnetization becomes much easier with the rotating frame. Figure 6 shows a series of situations when different RF pulses are
applied to the spin system placed
in Bo field. In the equilibrium state (a), the
net magnetization along the z-axis has its maximum value, M, = MI].The net
magnetization vector can be flipped by an angle a from the z-axis if individual
x-y
nuclei absorbed RF energy and shifted to the higher energy state (b). An
Fig. 5 Netmagnetizationin a rotatingframe.
Ruan and Chen
170
y = o
z'
%= "0
Fig. 6 Interactions between RF pulses andnet magnetization.
plane component is introduced by the RF pulse with a value of M,, = MI, cos
a, while z'-axis net magnetization is M, = M,, sina. It can be seen that flipping
M, and increases Mxy.The net
the net magnetization from the z'-axis reduces
magnetization can be flipped to any angle by application of a suitable RF pulse.
The 90", 180", and 270" RF pulses in the x'-direction, for example, will rotate
the net magnetization to the ??'-axis, z'-axis, and $-axis, respectively. After the
application of an RF pulse, the net magnetization will return to its equilibrium
state. During the return, the net magnetization along the $-axis and on the x'-y'
plane changes with time, in the form of recovery or decay. It is a process where
the excited spins release their energy and go back to the stable state or a process
in which the excited spins interact with each other and lose their phase coherence.
These processes are termed relaxation processes, which will be discussed next.
The ability to flip the net magnetization through RF pulses allows us to
manipulate magnetization so that the characteristicsof the nuclei can be observed
and detected. Sometimes, a sequence of RF pulses is used to generate the magnein this chapter.
tization signal. We will discuss the topic later
C.
171
RelaxationProcesses
After excitation, the spins return to their equilibrium states or equilibrium population distribution during the periodof free procession. The majority upward transition population, that is, the originally lower level population, returnsits to
equilibriumbylosingenergy
in theform ofan RF waveviavariousradiationless
transition processes called "relaxation processes."The RF wave signal
is characterized by the Larmor frequency of the nucleus and can be received and recorded
by the NMR instrument (RF-receiving antenna).
of a
The magnitude of M, and M,, versus the time after the application
90" pulse is plotted in Fig. 7. The magnetization along the z-axis, M,, indicates
a recovery or growth of the signal, which was first flipped to the y-axis direction
with M, = 0 by the 90" pulse. The recovery of M, is in exponential form and
eventually equilibrates to Mo. The x-y plane magnetization, M,,, is at its maximum value (M,,) immediately after the 90" pulse, after which it decays to 0.
The recorded signal of recovery or decay is characteristic of the relaxation
processes. There are two kinds of relaxation processes: spin-lattice (or longitudinal) relaxation and spin-spin (or transverse) relaxation.
The time constants describe these exponential relaxation processes and are known as relaxation times.
"_
Mo
<
nz
. .
Fig. 7 Relaxation of NMR signals after a 90" RF pulse.
Ruan and Chen
172
The spin-lattice relaxation time is designated by T I , and the spin-spin relaxation

time is designated by T2. Relaxation time is a function of the spin species and
the chemical and physical environments surrounding the spins.
In other words,
of the chemical and
the relaxation time constants are a fundamental property
physical environment. Therefore, analysis of T I and T2 of a sample will allow
the study of the chemical and physical properties of the sample. A long T I or T?
indicates a slow relaxation; a short
T , or Tz value indicates a rapid relaxation.
The relaxation time constant should not be confused with relaxation
rate. The
relationship between relaxation rate R and the time constant is simple and takes
the form of
Rl
T,
where I takes the value of 1 and 2 for spin-lattice relaxation and spin-spin relaxation, respectively.
Figure 7 also shows that M, reaches its equilibrium after approximately
five Tls, suggesting that the time to reach complete relaxation should be greater
than five times the spin-lattice relaxation time constant. This
is important in NMR
experiments, for which complete relaxation is often required.
1. Spin-LatticeRelaxation, T,
The lattice here means the environment surrounding the nucleus, including the
remainder of that molecule and other solute and solvent molecules. The energy
of the resonant nucleus is transferred to the various components of the lattice as
additional rotational, transnational, or vibrational energy until thermal equilibrium is reached between the nuclear spin and the lattice. The more the lattice
components rotate at the resonance frequency, the more
efficient or faster the
spin-lattice relaxation. In other words, there is a distribution of rotation frequenTI. Because
cies in a sample of molecules. Only the Larmor frequency affects
of
the Larmor frequency is proportional to B,,, T I will also vary as a function
magnetic field strength. In general, T I is inversely proportional to the number of
molecular motions at the Larmor frequency.
2. Spin-SpinRelaxation, T2
Chemical exchange or mutual exchange of spin states by two neighboring spins
is the typical form of spin-spin relaxation, although the sources contributing to
spin-lattice relaxation also lead to spin-spin relaxation. Unlike spin-lattice relaxation, spin-spin relaxationis an adiabatic process, and the redistribution
of energy
among thc spins doesnot change the numberof nuclei i n the higher energy level.
173
3. RelaxationMechanisms
Excited spins may reach equilibrium state with their environment (lattice)
in varito equilibrium:
ous ways. The following sources can cause a spin system to come
Dipole-dipole coupling caused by magnetic moments of other nuclei
Paratnagnetic relaxation caused by magnetic moments of unpaired elcctrons
Anistropic electronic shielding due to angular variations i n the electronic
shielding of the stationary B,, magnetic tield
Electric quadrupole relaxation causcdby electric quadruple moment of the
nucleus being irradiated
Spin-rotation relaxation as a result of molecular magnetic monuments
for many
Dipole-dipolecoupling is thedominantrelaxationmechanism
spin systems and involves neighboring spins (magnetic dipoles). This interaction
is referred to as dipole-dipole coupling because the motionof one spin is coupled
into the motion of the other. Each of these spins is affected not only by the B,,
tield but also by the magnetic moment of the other; the latter is dependent on
the magnetic moments of the two spins and the distance and relative orientation
between the two spins. For more details about rclaxation mechanisms, the reader
is referred to the texts of James and McDonald (2), Field (3). and Farrar (4).
D. RelaxationTimesandMolecularMobility
Both T , and Tz are related to molecular mobility in a system. T2 is proportional
to the rate of rotational motion of thc molecules. As a general rule,Tz is inversely
proportional to the molecular weight. A larger T2 generally means greater mobility. As we mentioned earlier, spin-spin relaxation is affected by a chemical exchange process. which is especially true in biological systems like foods. Chemicalexchangetakesplacebetweendifferentsites
of differentmobilities.For
example, protonso f water exchangewith protons ofmacromolecules. and protons
of less mobile water
of more mobile water molecules exchange with protons
molecules. etc. The rate of exchange has great influence on the T? value. If the
exchange is very fast, the differencein T? of the two exchanging sitesis averaged.
On the other hand, if the exchange rate is low, a broad distribution of T2 and
even two separate Tz sites can be expected. The spin-lattice relaxation process
is slightly different, although in normal situations a larger T i also means greater
mobility. For very fast motion, T i approximates T?. If the motion is extremely
slow (e&, at very low temperature or whenthe system is very rigid), T , increases
with decreasing mobility whereas T2 continues to decrease. Therefore, rigid or
solid-like materials could havea very short T? but a very long T i . We will discuss
Ruan and Chen
174
this later in this chapter. It appears that T2 is a more convenient indicator of

molecular mobility of a system.
While the proton intensity is related to the concentration
of proton-containing compounds in foods, relaxation times provide extremely useful information about the motional properties of food systems. Many researchers have succeeded in relating relaxation times to mobility of water, physical state of food
of foods, etc.
polymers, physiological state of plant origin foods, temperature
This presents a great opportunity for us to probe the quality and stability of food
products.
E. Measurement of Relaxation Times

Measurement of T, and T2can be accomplished through various pulse sequences.
The signals, normally the intensities, are recorded after one
or more pulses are
applied to the spins. The timingof each step during the pulse sequenceis of great
importance to the correct data acquisition and meaningful analysis and explanation of the data.
1. Measurement of T,
Because magnetization along the z-axis is difficult to detect, T, measurement is
impossible by using a single 90" RF pulse. Instead, the common approach is to
use an inversion-recovery pulse sequence
to generate the detectable signal.
In
RF pulse is followed by a 90"
the inversion-recovery pulse sequence, a 180"
RF pulse. The 180" pulse inverts the magnetization
to the -z-direction while
maintaining a magnitude -Mo. If the 90" pulse is applied immediately after the
180" pulse, the detected magnetization, M,,, will be equal to Mo. If we apply
the 90" pulse at some time T after the 180" pulse, the net magnetization has a
chance to decay somewhat due to the T , relaxation. Thus, the detected signal
along the y-axis will be slightly smaller than M,,. If this process is repeated with
longer T values, the delay times between the 180" and 90" pulses, the relationship
between the TI relaxation and time can be established through the following equation:
M,(z) = M(,(1
2e-T"'l)
(5)
Therefore T , canbedetermined by runningmultiple 180"- T-90"sequences

to equilibrium is
with varying T. A plot of M, at various times during return
shown in Fig. 8.
2. Measurement of T2
Tz describes the decay of M,, and is readily detectable. A 90" RF pulse flips the
magnetizationalongthey-axiswithamagnitude
of M,, = M,). M,, starts to
175
....................................................................................
Time
"0
r...................................................................................
Fig. 8 Change in netmagnetizationalongz-axisobservedduring

a 180"-t-90 pulse
sequence experiment. The recovery curve is in exponential form, from which T, canbe
determined.
shrink following the application of the 90' pulse in exponential form as mentioned earlier (Fig. 7). The decay curve is normally termed free induction decay
or FID and can be described by the following equation:
M,,
M,,e -'IT?
Therefore, I /Tz is the rate constant of the decay of a magnetization caused by a

90" RF pulse.However,unexpectedresultsareoftenobtainedwhenasingle
90" pulse experiment is performed. After the application of a 90" pulse, the net
magnetization decays more rapidly thanEq. (6) predicts. The spin-spin relaxation
time obtained from such an experiment is much less than the T2 defined by Eq.
(6) and is termed apparent T2 orTZ This has something to do with the magnetic
field inhomogeneity, diffusion, etc.
in different
If themagnetic field is notperfectlyhomogeneous,nuclei
of the field and hence
parts of the sample experience slightly different values
in
rotate at slightlydifferentfrequencies.Thiscanbeeasilydemonstrated
Fig. 9, which shows spins in three regions (I, 11, and 111) of the sample in the
magnetic field. In the case where the magnetic field is perfectly homogeneous
(Fig. 9A), all spins in the three parts of the sample are flipped to the x-y plane
along the y-axis as shown at time 0. The magnetization vectors M,, in the three
to MI).Because the magnetic field
parts of the sample are the same and equal
Ruan and Chen
176
RcgonII
RcgonI
Time0 +y-
Time A +y+.y+y+-y
... ....
TlmcB
....
Regon111
TotalNct
$-.$X
... ....
. . . ..
(A)
... .....
y T : y
RcgonI
y"
.....
RegonIII
$-.
.+y+.Y+>........
+y
.....
Region11
+y
TotalNeI
........
....
...
(B)
Fig. 9 Decay of x-y plane magnetization M,, in perfect ( A ) and imperfect (B)magnetic
field.
is perfectly homogeneous, all spins are rotating at the same frequency and pointof therelaxationprocess,although
ing in thesamedirectionatanytime
the net magnetization vectors shrink at later times because of T? relaxation. In
suchconditions i t is said that all the spinsare in phase orhavemaintained
phase coherence. In this condition, the net magnetization of the entire spin system at anygiventime
is equal t o the net magnetization of anyspins in the
system and can be described by the true
Tz. On the other hand. when the magin the
netic field is imperfect(Fig. 9B), thex-yplanemagnetizationvectors
to M,, afterthe 90" pulse,shrink by thesame
threeparts,initiallyequal
in slightlydifferentdirections
at
amountdue to the Tz relaxation.butpoint
times greater than 0 because the spins i n the three parts of the sample are rotating at different frequencies. This condition is termed dephase. and the spins
in thethreeparts
aresaid to losetheirphasecoherence.Becausethevectors
do not point in the same direction, the total net magnetization vector is shorter
thanthe total fortheperfectmagnetsystemthesametimeafterthe
initial
pulse.
The characteristics of such relaxation in an inhomogeneous magnetic field
is determined by TT, which includes contribution from natural molecular interactions (T?) and magnetic field inhomogeneity-related component (T2,,J:
177
However,because it isimpossibletodetermine
T?,,,withasingle 90" pulse,
90" pulse have been
many methods involving additional pulses after the initial
is to rephrase
developed for measurement of true T?. The common approach
or refocus the spin magnetization some time after the
initial pulse. Spin-echo
pulsesequence ( 5 ) is acommonmethodforT2measurement.Thespin-echo
sequence consists of two pulses, a 90" and a 180" pulse, in the form of 90"7-180" (Fig. 10). Figure 1 I illustrates the principle of spin-echo method. When
the spins are placed in the stationary magnetic field Bo, all the spins align with
the B,,-direction or z-direction (Fig. 1 IA). After the application of a 90" pulse,
to M,, (Fig.
all the spins are aligned along the v-axis, giving a nonzero value
1 le). M,, commences to decay because of T? relaxation and inhomogeneity
in themagnetic field. The field inhomogeneitycausestheindividualspins
to
in alossofphasecoherence,as
precess in differentfrequencies,resulting
in the x'-y' plane, as some spins move
discussed earlier, and they will fan out
faster (f) and some slower (s) (Fig. 1 IC). If after a time T a 180" pulse is applied,
the spins will be flipped 180" about the x'-axis, as shown in Fig. 11D. Because
theindividualspinsare
still precessingatthesameindividualfrequenciesas
before, the spins will
all be in phase again at time 2~ with the orientation in
the -v'-direction (Fig. 1 1 E). This is called refocusing or rephrasing. The continuing processing of the spins causes the spins again to lose their phase coherence
(Fig. 1 IF).
The rephrasing of the spins causes a free induction signal to build
to a
maximum (i.e., refocus) at time 2 T. This signal is termed spin-echo or
SE ampli-
90"RF
180"RF
Pulse
Pulse
Fig. 10 The spin-echoexperiment (9O"-t-18O0) induces a peak alnplitude or "echo"

at time 2 t .
"-c
178
Ruan and Chen
90" pulse
Y'
X'
03)
180" pulse
Dephase
X'
(C>
Rephase
X'
(E>
Fig. 11
X'
(F>
A simple spin-echo experiment: 90"-r-180".
tude. However, the dephasing process occurs not only because of the field inhomogeneity but also involves decay due to the natural transverse relaxation with
a time constant T2.If the 180" pulses are applied at different times, decay in SE
amplitude will be observed, characterizedby T2. Therefore,T2can be determined
by carrying out a separate pulse sequence for each value of t and detect the SE
amplitude at time 2 2 . However, the spin-echo method is limited in its range of
applicability because of the effect of molecular diffusion. The precise refocusing
of all the spin magnetic moments is dependent upon each nucleus remaining in a
constant magnetic field during the timeof refocusing experiment ( 2 ~ )If. diffusion
causes nuclei to move from one part of an inhomogeneous field to another, the
echo amplitude is reduced.
179
Fig. 12 CPMG experiment (90"-z- I80O-3~-180-5r-l80"-. . .) produces a train of echoes

at time 2 2 , 4 ~62.
. . . . The envelope ofthe multiple echo amplitudes reflects thc transverse
relaxation or decay, from which Tz canbe measured.
Carr and Purcell (6) proposed a variation of the simple Hahln's spin-echo
method, which reduces drastically the effect of diffusion. Their method involves
a 90" pulse at time t = 0 followed by a series of 180" pulses at times t, 3'1, 5t,
. . . , which refocus the individual spins to form echoes at times 2t, 4.t. 6t. . . .
However, the inaccuracy in the 180" pulse width can cause some errors in the
as the
spin-echo magnitude, which can be cumulative, becoming more serious
number of 180" pulses in the Can-Purcell sequence increases. The Carr-Purcell
sequence was modified by Meiboom and Gill to eliminate the errors. The CarrPurcell-Meiboorn-Gill or CPMG pulse sequence is the most common method for
Tz measurement. It uses the same pulse sequence as the Cam-Purcell technique,
but the 180" pulses are applied along the positive y'-axis. Thus, all of the subsequent refocusing is along the y'-axis, and a l l of the echoes are positive (Fig. 12).
The drawback of the CPMG experiment is that it takes a long time and is
unable to detect spins that relax so fast that their signal ceases to exist before
CPMG begins data acquisition. Therefore,the one-pulse (90") experiment is still
useful in obtaining information about the fast decay typically solid component.
to magnetic
For solid samples, T2 is so short that the apparent relaxation due
field inhomogeneityI/T?,,, is unimportant,and 1/TT = 1 /Tz + I/T,,,, 1/T2, or
Tz J: T?:.
F. MRI Methodology
MRI is an extension of NMR. It provides additional spatial information regarding
spins.The first publishedMRimage is credited to PaulLauterbur (7) in the
Chen
180
and
Ruan
early 1970s. The image was produced

using projection-construction computer
algorithms borrowed from computed tomography (CT) scanning. This method is
often referred to as backprojection reconstruction imaging. Because this method
requires a large number of "snapshots" at many angles covering
an arc of at
least 180" to obtain a cross-sectional image, modern MRI techniques have chosen
other more efficient methods. However, the key idea, which will be discussed
later, is still to superimpose magnetic field gradients onto the mainmagnetic
field, except that modern MRI techniques not only make the resonance frequency
a function of the spatial origin of the signal but also relate phase to the spatial
origin of the signal and expand the magnetic field gradients for many other uses
in the MRI image acquisition.
The MRI system is designed to excite and receive signals from a single
point, a line, a plane, or a three-dimensional volume in a sequential manner. The
point and line scanning methods are inefficient and less useful
and have been
superseded by two- and three-dimensional methods, which are more efficient.
The two- and three-dimensional methods apply linear field gradients to provide
spatial information.
1. Magnetic Field Gradient and Frequency Encoding

From Eq. (1) we know that the Larmor frequency
of a spin is proportional to
the magnetic field that it is experiencing,
If three tubes of water are placed in the same field, there will be only one
frequency and one peak in the NMR spectrum (Fig. 13). If a linear field gradient,
G, is superimposed on the main magnetic field, Bo, water tubes in different horizontal locations will experience different field strength because the field strength
varies with horizontal location in this case. Therefore, there will be two frequencies, and two peaks will appear in the NMR spectrum. The amplitude of the
signal is proportional to the number of water tubes or spin density.
Because Larmor frequency is proportional to the field strength [Eq. ( l ) ]
and field strength is a functionof position, we can establish a relationship between
frequency and spatial position of the spins by an equation similar
to Eq. (1):
r(Bo
+ xG,)
= Wo
+~xG,
(8)
or
where x is the position in x-axis direction and G, is the field gradient applied in
the x-axis direction. Here we say the frequency is spatially encoded. Field gradients can be appliedin all three directions. The symbols for a magnetic
field gradient i n the y- and z-directions are G, and G,. Similarly we have
181
S
I
Fig. 13 Application of a field gradient can spatially encode the signals from different
spatial positions. The three black spots represent three tubes of water in the sample.
Equation (8) forms the basisfor all magnetic resonance imaging methods.
The backprojection reconstruction imaging methodwill be presented in thenext
section to illustrate how the MR image is produced.
2. ProjectionReconstructionImaging
This method is of great historical importance since the first MR image was pro(7). Backprojection reconstruction is an extension
duced by Lauterbur in this way
of the frequency encoding discussed above. A one-dimensional linear
field gradient is applied at many different angles covering an arc of at least 180" and up
to 359". An NMR spectrum is recorded at each angle (Fig. 14A). Once a set of
data are recorded, the data are backprojected through space using a computer
program, and an image is reconstructed (Fig. 14B).
Chen
182
and
Ruan
Fig. 14 Principle of backprojection reconstruction imaging.(A) A one-dimensional field

gradient is applied at slightly different angles covering an arcof I 80, and NMR spectrum
is recorded. (B) A cross-sectional image is reconstructed by backprojecting the signals
through space.
3. FourierTransformImaging
Fourier transform (FT) imaging is currently the most widely usedMRI technique.
in 1974 by
The first two-dimensionalFouriertransformimagewasproposed
Kumer et al. (8). In FT imaging, both amplitude information and phase information are acquired to spatially encode the signal. The amplitude information (frequency encoding) is obtained by using the originalfield gradient, G,, from which
x can be calculated. Phase encoding is achieved by applying a phase encoding
gradient just before the originalfield gradient is turned on, and at the right angle
to the original gradient, normally in y-direction, denoted by G , or Go. The phase
encoding gradient is raised from zero in a series of small successive steps (time
duration t). Phase angle Cp is given by:
Cp
= 'Iy
1:
G,(t)dt
(12)
Due to the application of G, for a period oft, a point at a given y position

is off resonance by an amount of AW = -yyG,, and in time t the magnetization
vectorrotatesat an angle Cp = A u t = -yyG,t. @ canbedeterminedthrough
-y, G), and t, we will be able to
experiment (9). Knowing the phase change,
G, will contain both the
calculate y. Data collected after the original gradient
n is the
frequency and phase encoded information, and the number of the data
set sampling points. If we increment G , and repeat the procedure m times, we
183
will get a data set containing tn rows and n columns. By using Fourier transform,
we can produce an t n X tz image.
4. SliceSelection
Data for production of MR images are collected from spins of a plane through
an object. This plane will have certain thickness and is therefore termed a slice.
An MRI instrument is capable of selecting a slice of certain spatial location and
thickness. Slice selectionis of great importance to signal-to-noise ratio and image
resolution. Slice selectionis achieved by applying a one-dimensional, linear magnetic field gradient during the period that the RF pulse is applied. A 90" pulse
applied in conjunction with a magnetic field gradient will rotate spins, which are
look
located in a slice or plane through the object. To picture what this would
to examine
like if we had a cube of small net magnetization vectors, we need
of frequencies.
the frequency content of a 90" pulse. A 90" pulse contains a band
This can be seen by employing the convolution theorem. The frequency content
of a square 90" pulseis shaped as a sinc pulse. The amplitude of thesinc function
is largest at the frequencyof the RF, which was turnedon and off. This frequency
will be rotated by 90", while other smaller and greater frequencies will be rotated
by lesser angles.
The application of this 90" pulse with a magnetic field gradient in the xdirection will rotate some of the spins in a plane perpendicular to the x-axis by
90". The word "some" was used because
some of the frequencies have a B ,
less than that required for a 90" rotation. As a consequence, the selected spins
do not actually constitute a slice.
5. Pulse Sequences
The basic procedure for productionof a two-dimensional MR image can be summarized into five stages:
I.
2.
3.
4.
Excitethespins of selectedslice
Apply a phase encoding gradient for
a fixed time
Apply a frequency encoding or read gradient and collect 12 data points
Increment the value of phase encoding gradient and repeat steps
I to
3 t n times
5. Perform two-dimensional Fourier transform on the data to produce an
n z X n image
Therefore, pulses are applied in a sequential order. A schematic diagram

of a
typical MRI pulse sequence is shown in Fig. 15.
Other MRI pulse sequences provide additional information. These include
the three-dimensional imaging sequences, multislice imaging sequences. oblique
imaging, spin-echo imaging, fast imaging sequences, gradient-recalled smallflip-
184
Ruan and Chen
RF
Slice Selection Gradient
Phase Encoding gradient
Frequency Encoding
or Read Gradient
Signal Detection
Fig. 15 A two-dimensional MR imagingsequence.
angle imaging, echo planar and hybrid imaging, etc. Due to limited space, these
techniques will not be discussed in this chapter.
6. Imaging of Relaxation Time, Chemical Shift, and

Flow Velocity
In addition to the spin density, other NMR parameters such
as relaxation time,
in MRI.
chemical shift (IO), and flow velocity can also be used as contrast agents
To use these additional parameters to construct MR images, some modifications
to the basic MRI sequence have to be made. These additional contrast agents are
strongly associated with the mobilityof protons, which is governed by the physical and chemical environments in which the protons are embedded. Therefore,
imaging of these parameters will provide valuable spatial information about the
physical and chemical characteristics of the samples.
Many have used T, mapping to determine distribution of temperature and
mobility in food systems. The commonly used T, imaging methods are inversion
recovery and spin-echo imaging sequences, both considered to be time consumTI by
ing. While the multipoint inversion-recovery method, which determines
fitting a recovery curve to a series of measurements on a pixel-by-pixel basis, is
NMR Analysis
in Food Quality
185
the most accurate, its typical measurement time

is more than a half hour since
The
it requires a long repetition time (TR) and varying inversion delay times.
spin-echo method involves a standard spin-echo sequence with two TR values
and takes less time but still requires minutes for data acquisition.
A fast T, imaging pulse sequence is being developed in our laboratory.
The basic idea is to eliminate the waiting periods within conventional imaging
sequences by employing RF excitation pulses with flip angles considerably less
than 90", e.g., 10". The RFpulse is applied in the presence of a gradient affecting
only a certain slice of spins out of the three-dimensional object. Applying phase
encoding and read-out gradients in the two directions achieves two-dimensional
spatial discrimination, respectively.The signal is detected in the formof gradient
echo generated by the reversal of the read gradient. To map T I , a series of images
are acquired during magnetization relaxation following a single inversion pulse,
almostwithoutdelaybetweentwoadjacentimages.This
is thekeypoint
to
TI image. Currently, this TI imaging
achieve very short acquisition time for a
TI mapping techniques and can
technique is 20-100 times faster than previous
be used for dynamic food processes during which temperature changes rapidly.
111.
APPLICATIONS OF NMR AND MRI TECHNIQUES IN

FOOD QUALITY ANALYSIS
The intensity and motional properties of protons are two of the most important
NMR properties from which a number of analytical methods have been developed. The abundant presence of protons in foods makes them excellent image
of
contrast agents, enabling nondestructive observation of the internal structure
food products. The initial signal intensity (before decay or relaxation) is directly
proportional to the number of nuclei present. Therefore, it is possible to measure
the moisture and fat contents of a food product by measuring its proton signal
intensity. Through MRI, mapping of moisture and fat in food products can be
achieved. Similarly, the relaxation time constants,
T I and Tz, can be related to
molecular mobility of food products, as discussed earlier, and canalso be mapped
out using MRI. Previous research has demonstrated that variations in relaxation
is interaction between liquid and
times arise from many reasons. Among those
solids in the system. The interaction between liquid and solids is governed primarilybymoisture content, composition, particle size, structure, temperature,
and physical state of the system. Therefore, relaxation times can
be used to probe
many aspects of food systems.
A.
InternalStructure of Food Products
MRI has sinceits invention mainly been used to noninvasively visualize the internal structure of the human body for diagnosis purposes. In recent years MRI has
Ruan and Chen
186
Fig. 16 Multislice MR images of fresh orange.
been increasingly applied to food systems.

MRI is now capable of producing
two- and three-dimensional images, as shown
in Figures 16-1 8. Figure 16 shows
In addition, the darker regions repreorange segments that are clearly identified.
sent the seeds, which had weaker signals probably due to less moisture content
and/or a more rigid structure. This also suggestsMRI
thatcan be used to examine
external objects such as plastic, wood, and metal pieces that
may be accidentally
17 shows that cooking weakened the signal
mixed with food products. Figure
Raw
Cooked
Fig. 17 MR images of eggs. The brighter the image, the stronger the signal.
187
Fig. 18 Three-dimensional MR image of corn on the cob.
intensity from egg, which may be attributed to the tighter structure and stronger
18 demonwater binding due to heat denaturalization of egg proteins. Figure
strates the possibility of using MRI to acquire three-dimensional images, although
the procedure for getting a three-dimensional image is very complicated.
B. Ripeness and Bruise of Fresh Plant Foods

MRI has proven to be an excellent tool for illustrating the differencesin N M R
properties of fruits and vegetables in different development stages. Ishida et al.
(11) studied the distribution and relaxation times of water in ripe tomato using
MRI. They found that waterwith a long relaxation time was preferentially accumulated in seeds and seed envelopes in immature green fruit. The total amount
of water increased in mature red fruit, where water with a long relaxation time
was localizedin the outer wallsof the pericarp and water
with a shorter relaxation
time was distributed throughout all tissues except for seeds and seed envelopes.
MRI was able to distinguish the physiological variations among different types
of tissues and the physiological changes during maturation of tomato fruit.
MRI has been used to evaluate the internal quality factors of fruits
and
vegetables (12-16). These quality factors include bruises, dry regions, worm
damage, stage of maturity, and presence of voids, seeds, and pits. Wang et al.
(14) used the MRI method to obtain images
of water core and its distributionin
Red Delicious apples. The uneven distributionof mobile water and itsTI and T2
shown in the MR images was attributed to the variations in internal structures
of fruits, including petal bundle, endocarp, outer limit of carpel, carpel dorsal
bundle, receptacle cortex, receptacle pith and seeds, and to the difference
in NMR
properties between the normal and water core-affected tissues. The MR images
indicated that water core occurred primarily in an area 20 mm from the center
Chen
188
and
Hard
Ruan
Bruised
Soft
Fig. 19 h4R images of unripe, ripe, and bruised strawberries.
of an affected fruit, and the area most affected was between 5 and 10 mm from
19 also shows that the bruised tissue had
the center toward the stem end. Figure
stronger signal intensity, probably due to increased water mobility in the bruised
tissue.
Figure 20 shows two MR images of kiwifruits, one unripe (left) and the
other ripe (right). These images are T2-weighted. Inwords,
other they have partial
information on the mobility of water because of the way the imaging was conducted. In this type of MR image, the high intensity is attributed to both the
amount and mobilityof protons in the sample. In the unripe,firm fruit (left), the
Raw
Fig. 20 M R images of unripe and ripe kiwifruit.
Mature
189
core has a very low intensity, suggesting that this part of tissue is very firm and
low in moisture content. In contrast, the tissue surrounding the core appears
very
bright, indicating a soft texture with alot of moisture. The signal intensity of the
As the fruit ripens, the
outer flesh layer is between the core and middle layer.
signal intensity in the fleshy portion increased dramatically while
the intensity
in the core remained low, forming a sharp contrast between the fleshy and core
tissues. The three images shown in Figure 19 are cross sections of strawberries.
The effect of ripening on the NMR signal intensity of strawberry is similar to
that in the kiwifruit.
C.
Distribution of Moisture and Fat: Sensory and

Microbiological Concerns
Water and fat have a profound impact on the sensory quality of food products.
Equally important in this matter is the distribution of water and fat throughout
the food product. Distribution of water and fat in foods may be uneven due to
improper processing, handling, and redistribution
of water and fat during proin variations in texture and chemicessing, storage, and transport. This may result
cal reactions, resultingin inhomogeneous product quality.In addition, since water
is essential to the microbiological activitiesin food products, uneven distribution
of water may cause deterioration in regions having higher moisture contents.
MRI is anidealtoolforinvestigating
the distribution of water and fat.
However, it is a challenging task to separate signals of water protons from those
of fat protons since protons in both the water and fat molecules contribute to the
NMR signal in a conventional proton imaging sequence. Differentiation between
moisture and fat is essential to the acquisition
of MR image of protons from
watermoleculesonlyandfromfatonly.(Thetermswater-onlyandfatonly will be used here to refer to MR images of protons from water molecules
only and from fat molecules only, respectively.) In medical applications, numerous techniques have been proposed to generate moisture-only and fat-only imof proton signal while supages. Basically, these techniques promote one type
pressing the other type of proton signal. The principle of these techniques lies
in the difference in relaxation times and/or resonance frequencies between water
and fat protons.
The protons in water and fat molecules are found in different physical and
chemical environments and, correspondingly, have slightly different resonance
frequencies. The difference in resonance frequency between water and fat, also
termed chemical shift, is about 3.5 ppm or about 700 Hz in a 4.7 T magnetic
field. Their relaxation times usually are not the same. Most current moisture and
fat separation techniques in MRI take advantage of the differences in resonance
frequency or relaxation times and can be classified into chemical shift select-
Ruan and Chen
190
ive (CHESS) presaturation, frequency-selective excitation, Dixon's method, and

relaxation-based technique. The CHESS is widely used in clinical area for its
ease to implement and reasonable accuracy.
Rosen et al. (17) used CHESS presaturation technique to image lipid and
water in a human forearm. Haase et al. (18) imaged water and lipid in a human
hand and presented the 'Hspectra obtainedby selective elimination of H 2 0 and
CH2. Keller et al. (19) suggested multisection fat/water imaging with CHESS
presaturation, and obtained water-only and fat-only images
of a right knee, a
of the upper abdomen
spine and a head. Semelkaal.et(20) acquired water images
by presaturation of fat to increase the depiction of abnormalities and minimize
loss of anatomic detail around the edges of structures surrounded by fat. Mao et
al. (21) obtained water images by suppressing fat with an improved selective
presaturation pulse. The improved pulse has a broad and
flat pass band and a
in medical studies, we have
sharp transition band. In addition to applications
used CHESS sequence to acquire water-only and fat-only images (Fig. 21) for
cheddar cheese blocks(22) with reasonable good suppression quality, as shown
in Fig. 22.
In the CHESS imaging sequence, the first soft 90"RF pulse is frequency
selective with carrier frequency set at the unwanted spin resonance frequency.
It flips the unwanted spin magnetization into the transverse plane, and subsequently the spoiling gradient dephases the transverse magnetization. This presatu-
(4
(b)
Fig. 21 MR images of cheese: (a) water-only and (b) fat-only images.
191
-Water Saturated
-Fat Saturated
-No
20 22 24
Saturation
18 10 1612
14
ppm
Fig. 22 NMR spectra of cheese. Bottom: No suppression was applied. Middle: Fat was
suppressed. Top: Water was suppressed.
ration leaves the spin system in a state where no net magnetization of unwanted
component is left while the wanted component remains entirely unaffectedin the
form of longitudinal magnetization. A conventional spin-echo imaging sequence
is then applied to obtain the desired proton image.
D. State ofWater and Food Product Shelf Life

It has been pointed out several times
that water is of great importance to the
quality and stability of food products. Studies of properties of water in foods
have a long and distinguished history. Before the recognition
of water activity,
a,$, the moisture content of a given product was thought to be the critical factor
In thelate1950s,
in thecontrol of microbialgrowthandchemicalreactions.
microbiologist W. J. Scott (23) suggested the water activity concept, believing
that it could provide a better and more reliable measure of the availability of
by manyfoodandbiological
water.Sincethen, a,? has beenwidelyaccepted
scientists and industries and has been used as a primary guideline for the safely
and quality control of food, biological, and pharmaceutical products (24-26).
Today. a,, is still a treasure for many industries and is used extensively to model
product shelf life.
In recent years, the theoretical and practical limitations
of 11., have been
voiced by many food scientists and technologists (27-29).
The measurement of
a,of a food system is based on the assumption that the food system
is in its
equilibrium state, i.e., the measured partial vapor pressure above the food system
is presumed the same as that of water within the food system. This assumption
may hold true for infinitely dilute systems, where diffusion rates of water mole-
192
Ruan and Chen
cules are high compared to the time scale of the thermodynamic measurement.
Unfortunately, most food systems usually are nonequilibrium systems and, during
processing and storage, are maintained or brought intoa state of thermodynamic
instability, a state perhaps better designated as pseudo-stable since the unstable state may last longer than the normal lifetime of the product. Therefore, a
concept rootedin equilibrium thermodynamics, such
as a,, cannot fully represent
in
most food systems and should only be
used to describe systems that exist
equilibrium with their aqueous environments (e.g., water vapor) (28). Some researchers have begun to use relative vapor pressure (RVP) instead
of a, because
a, is in fact indicative of the rate at which water migrates out of the system.
Another defect of the aH.concept is that when different solutes are used to lower
the a, of a system, the response of the system in terms of the reaction rates or
stability is different. This suggests that measurement of a, is insensitive to the
solute-solute and solute-water interactions, whichin fact have a profound impact
on the reaction kinetics of the system.
The availability of water isa manifestation of how freely water molecules can participate in reactions or how easily water molecules diffuse to the
reaction sites to participate
in the reactions. The availability of water can be
reduced by adding certain chemical compounds, such as sugars (30,31). The reduced availability of water has been attributed to the increased viscosity of the
systems (3 1-34) or increased hydrogen bonding between water
and sugar molecules (35). Many researchers have found that the mobilityof water, as measured
by N M R and other techniques, is related to the availability of water in complex
systems (31,351-41). The higher the mobilityof water, the higher the availability
of water. Very mobile water molecules take
a long time to reach their equilibrium
state, or relax very slowly, thus having a small relaxation rate( R , or R?) or long
relaxation time (T, or T2, TI= 1/R, and TI = 1/R2). Figure 23 shows that added
sucrose suppresses the spin-spin relaxation time constant
of sucrose solutions.
N M R has been proven to be one of the most successful techniques to measure
the mobility of water, hence the availability of water (31,35-41).
1. Discrete and Continuum Models for Analysis of Relaxation

Times
Interpretation of N M R relaxation measurements has been model dependent. Because of the complex or heterogeneous characteristicsof food systems, many use
multiexponential models to analyze theN M R relaxation data. Mostof these models predicted relaxation decay data based on specific model assumptions, e.g., a
certain small integral number
of discrete exponential decay components
of differit may not produce
ent mobility (41,42). This does simplify the analysis, but
satisfactory solutions tovery complex, heterogeneous systems because
the factors
affecting spin-lattice and spin-spin relaxation behaviors are not yet well under-
193

2100
1700
1300
900
500
100
0
3010
20
40
50
60
Sugar concentration (Ye)
Fig. 23 Spin-spin relaxation time constant (TI) decreases as sugar content increases in
a sugar (sucrose) solution.
stood. In a heterogeneous system, spins exist in a wide variety of different environments, giving rise to a spectrum of relaxation times. In addition, chemical
and diffusive exchange, an important factor affecting spin-spin relaxation proa variety of T2 values, assuming there is a wide
cess, would also give rise to
range of exchange rates within the heterogeneous system (43-45). Therefore, the
measured relaxation decay is a sum of contributions from all spins, which have
sampled many different environments, or exchange with other spins at different
rates during the course of NMR experiment (46). It is thus reasonable to assume
that a continuum of relaxation times would arise from a continuum of different
exchange rates and different environments
in which spins exist. Some researchers
in various
have used the continuum models to follow the relaxation behavior
systems, although they focused on the state of water instead of that of biopolyal. (44) claimed that the
mers that interacted with water (44,46-51). Lillford et
continuous distribution of relaxation times is a better representation of the information content of the relaxation experiments.
The continuum approach seeks
a continuous distribution of relaxation times
and effectively adjusts a continuous variable number of degrees of freedom to
the minimum value necessary for a given data set. The CONTIN computer program of Provencher (52,53) has been used by researchers to process noisy data
Chen194
and
Ruan
including NMR relaxationdecaydata (S4,SS). It canproducerelaxationtime

(e.g., Tz) spectra that could be regarded as a probability distribution of spins at
various relaxation times. One of the advantages of the continuum approach is
that it conforms to the continuum natureof food systems. Furthermore, additional
information may be obtained from the continuum models. There are often several
a larger number of and/or broader
peaks on a T2 spectrum. A spectrum with
peaks would be expected for the heterogeneous samples but not for the more
homogeneoussamples.Therefore,thenumberanddegree
of variation of the
peaks could be used to indicate the homogeneity of the sample under analysis.
Proton relaxation is normally in exponential form, and the relaxation time
constants can be determined from the decay curves.For a 90 pulse, the resultant
free induction decay (FID) curve can be expressed as
where A is the amplitude at delay time t and AI, is the amplitude at equilibrium.
Usually, for heterogeneous systems like wheat flour dough. a multicomponent model (56) is used:
Equations ( 13) and (14) work well with simple systems. Both models have
been developed based on several assumptions, such as some predetermined number of discrete components in the decay curve (41).
WhittallandMackay (47) proposedamethodcallednonnegativeleastsquares (NNLS), which uses a continuum approach to analyze this type of data.
The general integral equation describing multiexponential relaxation is
where y(t,) is the observed amplitude measured

at time t, and S(T) is the unknown
amplitude of the spectral component at relaxation time T, which could be T, or
T?. The limits a and b are chosen to contain
the values of T expected for the
physical system being analyzed.
A linear inverse method has been developedto solve this equation. Assuming a large number of known relaxation times T, solved from the corresponding
amplitudes Sj. it is assumed that the spectrum is a sum of M delta functions with
unknown areas S, at known relaxation times T,:
J=
Substitution of Eq. (15) into Eq. (14) results in:
This is an equation of linear systems, whose general form
is:
,=I
where A,, is a matrix with elements exp(t,/TJ), and M is set large enough SO as
not to bias the solution into small number of relaxation times. Because of the
noise contamination, Eq. 18 cannot be solved exactly.
A non-negative least-squares algorithm has been developed for
this kind
of relaxation time analysis. in which extra constrains are incorporated into the
matrix A of Eq. 18 to alter the discrete character of the basic least-squares solution. A general form is to minimize
for fixed p > 0. a trade-off parameter. H, is a matrix representing K additional

constrains. and f k is the corresponding vectorof right-hand side values. The leastsquares solution is obtained when p = 0 (47).
CONTIN is a Fortran program for inverting noisy linear operator equations.
This program uses the NNLS method but, even more, is a general-purpose constrained regularization method, which finds the simplest solution that
is consistent
with prior knowledge and the experimental data. CONTIN has been proved as
a favorable approach compared with conventional NNLS and linear programming
approaches (57) or the Pade-Laplace method (54). The degree of successfulness
of this program is dependent on the number of temporal data points, the time
range of the measured data, and the signal-to-noise ratio (5435).
2. Mono-Exponential Decay
Proton relaxation is normally in exponential form, and the relaxation time con1I.E that
stants can be determined from the decay curves. We know from Sec.
T I can be determined by running an inversion recovery or saturation recovery
experiment, and T2 by a 90" pulse or a CPMG pulse experiment. Take the 90"
196
Ruan and Chen
t
Fig. 24 Freeinductiondecay (FID) curve.
pulse experiment, for example; the resultant FID curve is normally represented
by:
In A(t) = InA.
1
-
T2
where A(t) is the amplitude at timet and A. is the amplitude at equilibrium. The
plots of A versus t and In A versus t are shown in Figs. 24 and 25.
3. Multiexponential Decay
Figure 25 is a straight line as expected from Eq. (20). However, in many cases,
the plot of In A(t) versus t is not a straight line, as shownin Figure 26, suggesting
t
Fig. 25 Semi-log plot of mono-exponentialFIDcurve.
197
Fig. 26 Semi-log plot of multiexponential FID curve.
that there are two or more groups of water molecules that relaxat different rates.
The relaxation curve cannot be described by using the mono-exponential model
[Eq. 141. Instead, the multicomponent model described by Eq. (15) must be folin isolated amplitudesand relaxation time constants.
lowed. This model can result
Figure 27 shows a three-component model.
Fig. 27 Schematic diagram of a three-component model for multiexponential behavior

of proton relaxation.
198
Ruan and Chen
45
Fig. 28 Continuousdistribution of spin-spin relaxation times determinedbythe 90"

pulse experiment as a function of moisture content. Peaks on eachcurve at moisture contents of 12, 18, and 23% are labeled as PI and P2 from left to right. Above moisture
content of 23%, the single peak on the curve is labeled as P2.
4. Continuum Model
The example given here is from
our study of state of water
in wheat flour dough.
90"
Dough samples of different moisture were prepared and measured using the
and CPMG pulse sequences. Analysis of the data obtained from the 90" pulse and
CPMG experiments using the CONTIN package resulted in spectra (continuous
distribution) of T2 (Figs. 28,29). The x-, y-, and z-axes are T2value, amplitude,
T2 spectra computed
and moisture content, respectively. Figure 28 shows the
from data obtained from the90" pulse experiment.At moisture contentsof 12%
to 28%, two peaks (PI and P2) appear on each spectrum. Water molecules covered by these two peaks can be regarded as two groups having distinctly different
1 to 66 ps, proton
mobility. Because theT2values of these two groups range from
signals falling into these two groups can be regarded as from the solids (proteins
and carbohydrates) and/or water molecules
very close to the solids. Below moisture content of 23%, the increase in the area of the second peak and decreasein
average T2of individual peaks could be attributed to the increased available binding sites of the swollen flour substrates as a result
of addition of water (58). The
disappearance of the first peak at moisture above 23% may be due to the same
199
NMR In Food Quality Analysis
45%
O.OEMO\
loo
403
,~,
6579
-c
..--,
Fig. 29 Continuous distribution of spin-spin relaxation time determined by the CPMG

experiment as a function of moisture content. At and above moisture content of 23%,
peaks on each curve are labeled as P3, P4, and P5 from left to right.
reasons explained earlier. The increasein both peak area and T2 above moisture
content of 23% suggests that at 23% moisture level, all the water-binding sites
two
on the flour solids have been hydrated, and the additional water would be
or three layers away from these binding sites and therefore exchange and relax
more slowly.
The CPMG experiment was intendedto detect proton signals having relatively longer spin-spin relaxation time than the one-pulse experiment. The analysis of the data obtained from the CPMG experiments indicated that at moisture
content below 18%, no signal was detected, suggesting the dry flour had little
mobile water. At and above the 18% level, there were one to three peaks (P3,
P4, and P5) on the spectra (Fig. 29). T2values shown in Fig. 29 range from lo2
to IO5 ps, suggesting that the detected signals were from water molecules with
relatively high mobility. The number and size of peaks increased and the mean
T2values shifted to the right (increasing
T2value), with moisture content increasing up to 28%. The appearance
of new peaks suggests that new physical and
of addition of
chemical environments were formed within the system as a result
Chen
200
Ruan and
water to the system. This coincides

with the beginning of dough formation at
the moisture level above 23%.
It is noticed that the moisture content
of the dough samples affects the
shape of the spectra. A broader distribution could indicate a greater variation in
the chemical and physical properties of the system. Calculationof the coefficient
of variation (c.v.) of individual peaks would thus provide information about the
homogeneity of the systems under analysis. The following equations were used
to compute the coefficients of variation of spin-spin relaxation times (59):
SD
c.v.% = y x 100
X
where T2 and S are spin-spin relaxation time and amplitude, respectively, SD is

=
T 2 S i / CSi istheweightedaverage
of T2.
standarddeviation of T2, and
The results are shown in Fig. 30.
Figure 30 shows that the coefficientof variation of T2s in the range of 580 ps (the two peaks shownin Fig. 28) remained almost constant as the moisture
content was increased, suggesting that the environments
with which the solid-
+Pl
+P2
"
P
3
+P4
+PS
10
20
30
40
50
Moisture content (YO)
Fig. 30 Coefficient of variation of spin-spin relaxation times determined using the continuum model (see Figs. 28 and 29 for labels of PI, P2, P3, P4, and P5).
201
like and tightly bound water protons were associated did not change very
much while the moisture content increased substantially.
The coefficients of variation of those longer T?s (the three peaks shown in
of 40%, which
Fig. 29) showa gradual and slow increase below moisture content
of
may be a result of gradual formation of the bicontinuous network structure
dough and uneven distribution
of water among the flour constituents, as discussed
earlier. Upon reaching a moisture contentof 40%, which is more than the normal
dough moisture level, the coefficients of variation rose sharply, suggesting that
theexcesswater mayhavecreatedveryinhomogeneousdoughmorphology,
which should be further investigated.
E. Measurement of Glass Transition Temperature

Glass transition, a term in polymer science to describe a process in which a rubbery material, if cooled below a critical temperature, turns into a glassy material
(60,61), has made frequent appearances in food science and technology publications in recent years (62-73). Food scientists and technologists are fascinated by
the idea that foods are essentially polymers and can therefore be treated like
synthetic polymer systems in various analyses. Connections between the glass
transition and structural and textural characteristics, and chemical and microbial
activity of foods have been made (65,67-69,74).
Unlike low molecular weight substances that can exist
in three states of
aggregates-solids,liquid,andgas-polymersexistonly
in solidandliquid
states because they will decompose before they are vaporized, suggesting that
to break free
polymers never reach a mobility high enough for the molecules
even if heated to a very high temperature.
In low molecular weight crystalline solid substances, individual molecules
sit at their respective positions with a little vibrational motion but without translational or Brownian motions. When the temperature increases, more and more
kinetic energy is added to the system and the individual molecules move vibrationally and more and more vigorously. When the vibrational motion increases
sufficiently to cross the energy barrier that holds the individual molecules in the
of their fixed positions,
equilibrium positions, the molecules start moving out
activating Brownian motion. Eventually, upon further heating, the molecules difis
fuse all over randomly, and the well-defined molecular arrangement pattern
lost. By now, the substance is melted and is in a liquid state.
What happens when a high molecular weight amorphous or partially crysall
talline polymer is heated gradually? A high molecular weight polymer has
the characteristics of a low molecular weight substance at room temperature, but
when heated the difference between the polymer and the low molecular weight
substance can be seen. A high molecular weight polymer has a number of chain
segments. When we increase the temperature, some segments within the long
Ruan and Chen
202
chain of the polymer molecule are first mobilized before the whole molecule
starts moving. Further heating causes the entire molecule to move and become
liquid. From this discussion we can differentiate two types of motions: internal
or segmental motion and external or molecular motion. Both motions can be
regarded as Brownian. At a certain temperature,
both motions are frozen; at
higher temperatures both motions can be activated. These two situations correspond to the solid state and liquid states, respectively. What happens
if the polymer is placed at a temperature where only the segmental motion
is activated while
the molecular motion is frozen? The activated segments correspond to the liquid
state, while the molecule as a whole, with the mobility frozen, corresponds to
the solid state. This state, which is indeed a mixture of liquid and solid, is called
the rubbery state. On further heating the polymer in the rubbery state becomes
a highly viscous liquid and starts flowing; this state is called the viscofluid state,
the transition taking place at the flow temperature Tf (Fig. 31).
Polymers can have two types of motions: segmental and molecular. How
can we relate these motions to the NMR properties? Can we determine the Tg
by detecting the motional characteristicsof the polymer? Does waterplay a role
here?
NMR measuresprotonrelaxationcharacteristics,
Asdiscussedearlier,
which can be related to the mobility
of the proton-containing molecules. In a
liquid state, a long relaxation time constant indicates high mobility and increases
with increasing temperature. When the system is moving toward a solid state,
for instance, due to suppressing of temperature, the relaxation processes behave
somewhat peculiarly. We have mentioned that the spin-lattice relaxation is a process of energy release by the excited spins to the lattice (the entire molecular
0Motion
not activated
Motion
activated
Fig. 31 A schematicpresentation of therelationship between states of materialand

motional characteristics.
203
system). To facilitate an efficient energy release by the excited spins, the lattice
In the solid state, the number of
has to oscillate at the resonant frequency a,,.
molecules oscillating or rotating at the resonant frequency olIis very small, suggesting a very inefficient spin-lattice relaxation. Therefore, the dynamic contribution to the T I decay is very small, causing TI to become very long again. On the
is very short due to the decrease
in static contribuother hand,in the solid state, Tz
tion to transverse-component dephasing.
The dynamic motion of spins (or molecules) can also be characterized by
a correlation time z,, which is, roughly speaking, the minimum time required for
the nuclear magnetic moment to rotate one radian (I/? 7t of a complete circle). In
general, T, for a nonviscous liquid is very short. With water, for instance, T, is
about 10"' s. On theotherhand, T, forsolids is verylong,about
s. Within
Z~ and the more quickly a perlimits, the slower the motion, the longer will be
Z, and relaxation time
turbed spin system will relax. The relationships between
constants can be described as follows:
where K = 3y2/160d h?yJ/r?is a constant that includes a number of nuclear

parameters and constants. These equations have been proven very useful for the
understanding of a single relaxation process dominant system. Figure32 is a plot
of Z( versus relaxation time constants.
A plot of relaxation time constants versus reciprocal of absolute temperature is generally in the form shown in Figure 33. It can be seen that curves in
Figures 32 and 33 share the same shapes. This is because that, within limits, Z,
is related to temperature. Their relationship can be described by
7, = T , , , ~ ~ . ~ < I ' L ' '
(25)
where E:,,, is the activation energy for rotational motion,

k the Boltzmann conz~,,a constant. We should also be aware
stant, T the absolute temperature, and
of o0q..This is especially the
that the relaxation time constants are a function
case of T , . Figure 34 shows that the corresponding T, and magnitude of T I minimum shift when OIlis varied. From Eq. (23). we know that T I minimum occurs
when c o , ) ~=~ 0.616. We have
Ruan and Chen
Fig. 32 A generalized schematic diagram showing the relationships between relaxation

time constants (TI and TJ and correlation time (q).
Fig. 33 NMR relaxation time constants (TI and Tz) as a function of reciprocal absolute
temperature ( 1 IT).
205
10000
+T1
(20M H z )
+T1(200
MHz)
0.01
'
'
0.0001
l.E-10
'
"""'
'
1 .E-09
'
"""'
1 .E-08
"""'
'
1 .E-07
'
'"'.rl
1 .E-06
Tc (s)
Fig. 34 Theoretical curves of correlationtime T, versusrelaxationtime constants

and Tz)at different resonance frequencies.
where fll is the frequency of the NMR equipment related to the static magnetic
field Bo, i.e.,
BecauseformostcommercialNMRanalyzers,
fo is in therange of 10-600
MHz, o,]
is in the range of 60-3800 MHz. Therefore, for T , minimum to occur,
z, (= 0.616/o11)has to be in the range of 10""-10~* s.
Another model for calculation of z,, the Debye-Stockes theory, relates T,
to molecular size (with radius of a), medium viscosity 11,and absolute temperature T:
zc = 4na3q/3kT
(28)
where k is theBoltzmannconstant.Thismodelpredictsthatcorrelationtime
increases with larger molecules, viscous solutions, and low temperature. In other
words, within certain limits, the relaxation time constants for large molecules,
viscous solutions, and low temperature systems are small. On a qualitative basis,
Eq. (28) predicts a linear dependence of T, on T / q for some liquids (75).
However,thesemodelsare
not alwaysapplicable to complexsystems.
Many synthetic polymers, which have more than one group on the main chain
andprobablymorethanonecorrelationtime,undergomultiplerelaxation
Chen
206
and
Ruan
processes (76). Themotion of eachgroup is differentandmaybecomethe

dominant process in a certain temperature range. This multiple relaxation proT, minima. Each minimum is related to the
cess is characterized by multiple
motional behavior of a specific group and sometimes,
if the chemical structure
of the polymer is known, can be assigned to a specific group. However, with
increasing molecular weight or length, relaxation processes become more complicated, and the distinguished minima may merge due to the overlapping of
T, minimum
motions contributed by individual groups to form a single, broad
(77). For most food systems, there are so many different compounds in the systems that overlapping of motions experienced by individual compounds may occur over a range of temperatures. The dominant relaxation processmay produce
a single T, minimum where the segmental motion of the dominant compound is
activated.
Figures 35 and 36 show the dependenceof T, on the temperature for amorphous maltodextrin of dextrose equivalent (DE) of 5 and 25 with 25% moisture,
respectively. A single, broadT, minimum is observed for DE 5 over the temperature range tested, while the cure for DE 25 exhibits an L shape.
Figure 37 is a plot of T2 versus temperature. T? changes very little when
temperature is lowered to a critical point from which the solid begins its rigid
lattice behavior. The onset of rigid lattice behavior is characterized by T2 =
z,, which is about IO s, a value found for many solid materials. The midpoint
1000
T,minimum
100 :
I
~
10
100
200
I
300
400
Temperature (K)
Fig. 35 Spin-latticc relaxation time constant (T,)as a function of absolute temperaturc
for maltodextrin samples (DE = 5; moisture content = 25%).

1000
loo0
-GE
207
n
W
1 0 0 ;:
100
10
150
200
250
300
350
400
Temperature (K)
Fig. 36 Spin-lattice relaxation time constants as a function of absolute temperature for
maltodextrin samples (DE = 25; moisture content = 29.2%).
100
n
rA
3.
Transition
point
10
100
200
300
400
Temperature (K)
Fig. 37 Spin-spinrelaxationtimeconstantsasafunctionofabsolutetemperature
maltodextrin samples (DE = 5 ; moisture content = 25%).
for
and
208
Ruan
Chen
of the T2 transition on the T2-temperature curves is often considered related to

the glass transition process. We used differential scanning calorimetry (DSC 7,
The Perkin-Elmer Corporation, 20 mg of sample size, S"C/min of heating rate,
midpoint T, calculated using the system built-in computer program) to measure
the T, of the material and found that the transition temperature determined from
the TI-temperature and T2-temperature curves are very close to the DSC determined T,. This further suggests that the transition phenomena observed with
NMR relaxation experiments are indeed associated with the glass transition process. We will provide more examples later to demonstrate the use
of NMR in
determination of the glass transition process in various food systems.
Glass transition is largely affected by the composition
of the system of
interests.Amongthecompositionalfactors,moisturecontentandmolecular
T, (69,78,79).
weight of the macromolecules are the major players governing
Water as a plasticizer is capable of mobilizing the solid matrix and increasing
the mobility of the system. Large molecules have less mobility, hence larger 'cc,
and require larger input of thermal energy to mobilize the structure. Water is a
plasticizer that can penetrate into the polymer matrix and establish polar attractive
forces between it and the chain segments. These attractive forces reduce the cohesiveforcesbetweenthepolymerchainsandincreasethesegmentalmobility,
thereby reducing the T, value. Figure 38 shows a shift of minimum point on the
T,-temperature curves to lower temperature as moisture content was increased
in the maltodextrin samples. Similarly, the T2-temperature curves (Fig.39) show
1000 c
t
.
8.9%
+10.4%
+12.8%
10
'
200
250
300
350
400
Temperature (K)
Fig. 38 Effect of moisture content on spin-lattice relaxation time constantTI for maltodextrin (DE = 5).
209
1.35
A 14.70%
1.25
-3
1.15
1.os
x18.30&
0.95 I
250 200 150
300
350
400
Temperature (K)
Fig. 39 Effect of moisture contenton spin-spin relaxation time constantTzfor maltodextrin (DE = 5 ) .
a shift of the transition point to the lower temperature as the moisture content
of the samples increased.
F. MRI MAPPING OF GLASS TRANSITION TEMPERATURE

Many foods and biological materials are heterogeneous. Such heterogeneity
may
be caused by the incompatibility of ingredients, poor mixing, or relocation of the
ingredients during processing and storage. For example, the crust on the surface
of many baked food products such
as a cake is due to the excess heating and
loss of moisture during baking. Therefore, there will be an uneven distribution
a variable
of the physical structure and moisture content within the cake, and
distribution of T, is expected. For baked goods, an uneven distribution could
mean that the textural properties of the products are not uniform. For other products where chemical and microbiological activities are key deteriorating factors
that are strongly influenced by T,, an uneven distribution of T, could be a major
challenge to the safety control, which
is usually based on a single average parameter, e.g., T,. A localized low T, can put the corresponding spot in a condition
well above the T,, allowing chemical and microbiological activities to take place
at this very spot.
Currently there is no study on the measurement of T, distribution reported
in the literature. The conventional techniques
do not have the capabilityof provid-
210
Ruan and Chen
ing spatial information about the thermal dynamic or dielectric properties of the
materials. On the other hand, MRI can generate spatial information about the
nuclear relaxation characteristicsof the material, which can be related to the glass
transition process as discussed earlier in this chapter. Therefore,
if we can estabL = 1,2) encoded with
lish a relationship between relaxation time constants (TL,
the spatial information and temperature (T), i.e.,
TL(X, Y) = f(T)
(29)
we will be able to produce a Tgmap.

To construct aT, map of a sample, we need to obtain a series of TL maps
the TIminima or Tz transition
of the sample at different temperatures, from which
points and corresponding temperatures for each pixel are computed using a curvefitting program, after which a map of temperatures corresponding to T
the
I minima or T2 transition points, that is, a Tgmap, is constructed. This procedure is
demonstrated in Fig. 40A and B.
Figure 40A shows that, as temperature increases, TIvalues of individual
pixels changed. The two samples with different moisture contents responded to
temperature differently. The analysis of the average TIvalue of each image as
a functionof temperature revealed that the minimum
T I value was found between
-5 and 6C for the sample with25% moisture content and between
13 and 23C
for the sample with 18% moisture content. The temperatures corresponding to
the T I minimum values agree well with the Tg determined by the pulsed NMR
and DSC methods.
M C = 25%
M C = 25%
MC 1Wo
M C = 18%
47C30C-5C
23C13C 6C
Fig. 40 (A) Tlmapsof maltodextrin (DE 5 ) samples differing in moisture content(MC)

obtained at different temperatures (darker = high T, value). (B) Tg map of maltodextnn
(DE 5) samples calculated and constructed based on
TImaps shown in (B) (darker =
higher temperature).
211
This procedure demonstrates that two samples placed side

by side can be
measured simultaneously. This is a great advantage of the MRI mapping technique for simultaneous determination of T,s of multicomponent foods.
IV.
FUTURE TRENDS
NMR and MRI have found use in other areas not mentioned specifically above.
of
NMR is widely used for rapid determination
of moisture and fat contents
oilseeds. Use of MRI to measure rheology of food products has also been
reported. Attempts were also made to combine MRI and mathematical modeling
techniques to study simultaneous heat and mass transfer.
Efforts are needed to address the relationships between NMR properties
and sensory quality attributesin food systems so that NMR can be used for rapid,
nondestructive, and noninvasive evaluation of food products. There are also
a
need to develop affordable NMR hardware and more sophisticated analysis software.
GLOSSARY
180"pulse: RF pulse designed to rotate the net magnetization vector 180" from
the static magnetic field.
90" pulse: RF pulse designed to rotate the net magnetization vector 90" from
the static magnetic field.
B,,: Conventional symbol for the main magnetic
field in MRI system.
Carr-Purcell (CP) sequence: Sequence of a 90" RF pulse followed by a repeated 180" RF pulse to produce a train of spin-echos.
Carr-Purcell-Meiboom-Gill (CPMG) sequence: Modified CP sequence with
90" phase shift in the rotating frame of reference between the 90" pulse
and the subsequent 180" pulse to reduce accumulating effects of imperfections in the 180" pulses.
Correlation time: The minimum time required for the molecule
to rotate one
radian.
Echo: A form of magnetic resonance signal from the refocusing of transverse
magnetization.
Equilibrium: Themagneticstate of anobjectthat is fullymagnetized by a
static magnetic field.
Fourier transform imaging: A mathematical procedure used in MR that converts a time-domain signal into a frequency- or spatial-domain signal or
image. It is analogous to the way that our ear distinguishes or separates
out separate sounds or frequencies from noise we hear. Our eyes do
not
212
Ruan and Chen
work this way. If we see a mixture of blue and yellow we see the color
green, not the original blue and yellow.
Free induction decay (FID): A form of magnetic resonance signal from the
decay of transverse magnetization.
Frequency encoding: Use of a magnetic field gradient to produce a range of
frequencies along the MR signalto provide information on spatial position.
Gradient: Amount and direction of the rate of change in space of some quantity
such as magnetic field strength.
Larmor frequency ( 0 ) : The frequency at whichmagneticresonancecanbe
excited.
Lattice: Environmentssurroundinganucleus
or spin.
Magnetic moment: Measure of the magnetic properties of an object or particle
that causes it to align with the static magnetic field.
Magnetization: A vector quantity measuring the net magnetic moment
of nuclear spins per volume of a sample.
NMR signal: Electromagnetic signal i n RF range produced by the precession
of the transverse magnetization of nuclear spins.
Precession: A rotational motion of a spinning body about an axis
of a vector
whose origin is fixed at the origin.
Pulse sequence: A series of RF pulses and/or magnetic field gradients applied
to a spin system to produce a signal representative of some property of the
spin system.
Radiofrequency (RF): Electromagnetic radiation lower in energy than infrared. RF is in the range of 10 to 100 MHz.
Relaxation rates and time constants: After excitation, the spins tend to return
to their equilibrium state at certain rate. This rate is called relaxation rate.
The reciprocal of relaxation rate
is relaxation time. There are two relaxation
processes: spin-lattice or longitudinalrelaxation. I t is the return of longitudinal magnetization to its equilibrium value after excitation through exchange of energy between the spins and the lattice with a characteristic
time constant termed spin-lattice relaxation time T I .The second is called
the spin-spin or transverse relaxation process. in which the transverse component of magnetization vector, which is at right angles to the static magloss of transnetic field. decays towards zero. The characteristic time for
verse magnetization to zero is termed spin-spin relaxation time
T:.
Spin or nuclearspin: The intrinsic magnetic momentum of an elementary particle such as a nucleus responsible for
the magnetic moment. A fundamental
property of matter responsible for MRl and NMR.
Spin-echo: Reappearance of an MRIsignalaftertheFIDhasdisappeared.
It is the result of the effective reversal of the dephasing of the nuclear
spins.
213
REFERENCES
I.
A Abraham. ThePrinciplesofMagneticResonance.London:OxfordUniversity
Press,1960.
2. TL James, GG McDonald. Measurement of the self-diffusion coefficient of each
component in a complex system using pulsed-gradient Fourier transform NMR.
J
MagnReson11:58-61,1973.
3. LD Field. Fundamental aspects of NMR spectroscopy. In: LD Field, s Sternhell,
eds. Analytical NMR. Chlchester: John Wiley
& Sons, 1989:5-39.
4. TC Famar. Pulsed NMR Spectroscopy. Madison. WI: The Farragut Press. 1989:211 .
5. EL Hahn. Spin echoes. Phys Rev 80:580-594, 1950.
6. HY Carr, EM Purcell. Phys Rev 94:630, 1954.
7. PC Lauterbur. Image formation by induced local interactions: examples employing
nuclear magnetic resonance. Nature 243: 190, 1973.
8. A Kumer, D Welti, RR Ernst. N M R Fourier zeugmatography. J MagnReson28:
69-83,1975.
9. CNChen, DI Hoult.Biomedicalmangeticresonancetechnology.Bristol:Adam
Hilger, 1989340.
10. WG Proctor, FC Yu.The dependenceof nuclear magnetic resonance frequency upon
chemical compound. Phys Rev 72:717, 1950.
1 I . N Ishida, T Kobayashi, M Koizumi, H Kano. H-NMR imaging of tomato fruits.
Agric Biol Chem 53(9):2363-2367, 1989.
12. P Chen, MJ McCathy, R Hauten. NMR for internal quality evaluation of fruits and
vegetables. Trans ASAE 32:1747, 1989.
13. CY Wang, WPC. Non-destructive detection of core breakdown in Bartlett pears
with nuclear magnetic resonance imaging. HortScience 24( I): 106, 1989.
Nondestructive detectionofwatercoreinapplewith
14. CYWang.WPC,MFaust.
nuclear magnetic resonance imaging. Sci Hort 35(3-4):227-234, 1988.
1s. B Zion, P Chen, MJ McCarthy. Nondestructive quality evaluation of fresh prunes
by NMR spectroscopy. J Sci Food Agric 67(4):423-429, 1995.
16. JL Maas. MM Millard, MJ Line, GJ Galletta, JL Maas, GJ Galletta. Nuclear magnetic resonance microimaging of strawberry fruit. Acta Hort 348:375-377, 1993.
17. BR Rosen. VJ Wedeen, TJ Brady. Selective saturation NMR imaging.
J Computer
Assisted Tomography 8:813. 1984.
18. A Haase. J Frahm. WH Nicke,DMatthaei.IH
NMR chemicalshiftselective
(CHESS) imaging. Phys MedBiol30:341.1985.
19. PJ Keller, WWJ Hunter, P Schmalbrock. Multisection fat-water imaging with chemical shift selective presaturation. Radiology 164539. 1987.
20. RC Semelka, W Chew, H Hricak, E Tomei. CB Higgins. Fat-saturation
MR imaging
of the upper abdomen. AJR 155:I I 1 I . 1990.
21. J Mao, J Gao, H Yan, JR Ballinger. Susceptibility artifact reduction in fat suppression. MRM 33582-587. 1995.
22. K Chang, RR Ruan, PL Chen,RG Fulcher, E Bastian. Moisture, fat and temperature
mapping of cheese block during cooling using MRI. Paper N0.956535 presented at
ASAE meeting, Chicago, 1995.
214
Ruan and Chen
23. WJ Scott. Water relations of food spoilage microorganisms. Advances Food Res 7:
83-127,1957.
24. JA Troller. Influence of water activity on microorganisms in foods. Food Technol
(5):76-83,1980.
25. S Schwimmer. Influence of water activity on enzyme reactivity and stability. Food
Technol (5):64-74, 1980.
26. JA Troller. Water activity and food quality. In: TM Hardman, ed. Water and Food
Quality. London: Elsevier Applied Science, 1989: 1-3 I .
27. F Frank. Hydration phenomena: an update and implications for food processing industry. In: H Levine, L Slade, eds. Water Relationships in Foods-Advances in the
1980s and Trends in the 1990s. New York: Plenum Press, I99 1 : 1- 19.
of food safety and quality? Trends Food
28. F Frank. Water activity: a credible measure
SciTechnol:68-72,1991.
29. L Slade, H Levine. Beyond water activity: recent advances based on an alternative
approach to assessment of food quality and safety. Crit Rev Food Sci Nutr 30: 115360,1991.
ORFennema, ed.FoodChemistry. New York:
30. RCLindsay.FoodAdditives.In:
MarcelDekker,1985:629-687.
31. H-M Lai, SJ Schmidt. Mobility of water in various sugar-water systems as studied
by oxygen-I7 NMR. Food Chem 4655-60, 1993.
32. MJ Tai. A Suggett, F Frank, S Ablett, PA Quickenden. Hydrogen of monosaccharides: a study by dielectric and nuclear magnetic relaxation. J Solut Chem 1 :131151,1972.
33. PS Belton. KM Wright. An 170 nuclear magnetic resonance relaxation-time study
of sucrose-water interaction. J Chem SOC Faraday Trans 1(82):451-456, 1972.
34. GW Padua. Water States Associated with Skim Milk Components as Determined
by NMR. Urbana, IL: University of Illinois, 1989.
35. HS Lai, SJ Schmidt. Water mobility and crstallization action of lactose-water systems by oxygen-17 and carbon-13 NMR. J Food Sci 55:1435-1440, 1990.
36. SJ Richardson, IC Baianu, MP Steinberg. Mobility of water in wheat flour suspensions as studiedby proton and oxygen- 17 nuclear magnetic resonance.J Agric Food
Chem 34:17-23, 1986.
37. SJ Richardson, IC Baianu,MPSteinberg.Mobilityofwaterinsucrosesolutions
determined by deuterium and oxygen-17 nuclear magnetic resonance.
J FoodSci
52:806-809,1987.
38. A Mora-Gutierrez, IC Gaianu. IH NMR relaxation and viscosity measurements on
solutions and suspensions of carbohydrates and starch from corn: the investigation
of carbohydrates hydration and stereochemical and aggregation effects in relation
to "0 and '>CNMR data for carbohydrate solutions. J Agric Food Chem 37: 14591465,1989.
39. L Kakalis, IC Baianu, TF Kumosinski. Oxygen-17 and proton nuclear magnetic relaxation measurements of soy protein hydration and protein-protein interactions in
solution. J Agric Food Chem 38:639-647, 1990.
40. BP Hills. Multinuclear NMR studies of water in solutions of simple carbohydrates.
I. Proton and deuterium relaxation. Molec Phys 72: 1099- I 121, 1991.
41. SJ Schmidt, H Lai. Use of NMR and MRI to study water relations i n foods. In: H
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60,
61.
215
Levine, L Slade, eds. Water Relationships in Foods.

New York: Plenum Press, 1990:
405-452.
HK Leung, JA Magnuson, BL Bruinsma. Pulsed nuclear magnetic resonance study
of water mobility in flour doughs. J Food Sci 44(5):1408-141 I , 1979.
JR Zimmerman, WE Brittin. Nuclear magnetic resonance studies in multiple phase
systems: lifetime of a water molecule in an adsorbing phase on silica gel.
J Phys
Chem61:1328-1333,1957.
PJ Lillford, AH Clark, DV Jones. Distribution of water in heterogeneous foods and
model systems. In: SP Rowland, ed. Water in Polymers. 1980:177-195.
PS Belton, BP Hills. The effect of diffusive exchange in heterogeneous systems on
NMR line shapes and relaxation processes. Molec Phys 61:999-1018, 1987.
RM Kroeker, RMHenkelman.AnalysisofbiologicalNMRrelaxationdatawith
continuous distribution of relaxation times. J Magn Reson 69:218-235, 1986.
KP Whittall, ALMacKay.QuantitativeinterpretationofNMRrelaxationdata.
J
Magn Reson 84: 134-152, 1989.
RS Menon, PS Allen. Application of continuous relaxation time distribution to the
fitting of data from model systems and excised tissue. J Magn Reson 86:214-227,
1991.
CD Araujo, AL MacKay, JRT Hailey, KP Whittall. H Le. Proton magnetic resonance
techniques for characterizationof water in wood: application to white spruce. Wood
Sci Techno1 26(2):101-113, 1992.
C Tellier, F Mariette, J Guillement, P Marchal. Evolution of water proton nuclear
magnetic relaxation during milk coagulation and syneresis: structural implications.
J Agric Food Chem 4l( 12):2259-2266, 1993.
CH Newcomb, SJ Graham, MJ Bronskill. Effects of nonlinear signal detection on
NMR relaxation time analysis. J Magn Reson 90:279-289, 1990.
SW Porvencher. A constrained regularization method for inverting data represented
by linear algebraic or integral equations. Comput Phys Commun 27:2 13-227, 1982.
SW Provencher. CONTIN: a general purpose constrained regularization program
for inverting noisy linear algebraic and integral equations. Comput Phys Comrnun
27:229-242,1982.
C Labadie, JH Lee, G Betek, CSJ Springer. Relaxograph imaging.
J Magn Reson
105:99-112,1994.
JH Lee. Magnetic Resonance Studies of Tissue 23 Na and 'H20signals. State University of New York, 1993.
R Ruan, PL Chen. Water in FoodandBiologicalMaterials:ANuclearMagnetic
Resonance Approach. Lancaster, PA: Technomic Publishing Inc, 1998.
K Overloop,L Van Gerven. NMR relaxation in adsorbedwater. J MagnReson
100(2):303-315,1992.
W Bushuk, V K Mehrotra. Studies of water bindingby differential thermal analysis.
11. Dough studies using the melting mode. Cereal Chem 54(2):320-325, 1977.
JL Devore. Probability and Statistics for Engineering. Monterey, CA:' BrookslCole
Publishing Co., 1982.
LH Sperling. Introduction to Physical Polymer Science, 1986.
VR Gowariker, NV Viswanathan, J Screedhar. Polymer Science.New York: Halsted
Press,1986.
Ruan and Chen
216
62. HC Troy, PF Sharp.aandbLactose in somemilkproducts. J Dairy Sci 13:140157,1930.

63. R Katz, TP Labuza. Effect of water activity on the sensory crispness and mechanical
deformation of snack food products. J Food Sci 46:403-409, 1981.
64. GW White, SH Cakebread. The glass state in certain sugar-containing food products.
J Food Technol 1:73-82, 1966.
65. YH Roos, M Karel, JL Kokini. Glass transitions in low moisture and frozen foods:
Effects on shelf life and quality. Food Technol SO( I1):95-108, 1996.
66. YH Roos. Effect of moisture on the thermal behavior of strawberries studied using
differential scanning calorimetry. J Food Sci 52:146-149, 1987.
67. H Lavine, L Slade. Apolymer physicochemical approach to the study of commercial
starch hydrolysis products (SHPs). Carbohydr Polym 6:213-244, 1986.
68. H Lavine, L Slade. A food polymer science approach to the study ofcryostabilization
technology. Conm Agric Food Chem 1.3 15-396, 1989.
69. L Slade, H Lavine. Glass transitions and water-food structure interactions.Adv Food
Nutr Res 38:103-269. 1995.
70.JMFlink.Structureandstructuretransitions
in driedcarbohydratesmaterials. In:
M Pelega, EB Bagley, eds. Physical Properties of Foods. Westport. CT: AVI Publishing Co., Inc., 1983:473-521.
7 I . JMV Blanshard, F Franks. Ice crystallization and its control in frozen food systems.
In: JMV Blanshard, P Lillford, eds. Food Structure and Behavior. Orlando, FL: Academic Press, Inc., 1987:51-65.
72.HLavine,
L Slade.Collapsephenomena-aunifyingconceptforinterpretingthe
behavior of low moisture foods. In: JMV Blanshard, JR Mitchell. eds. Food Structure-Its creation and Evaluation. London: Butterworths, 1988: 149180.
73. H Lavine. L Slade. Influence ofthe glassy and rubbery states on the thermal, mechanical, and structural properties of doughs and baked products. In: H Faridi. JM Faubion, eds. Dough rheology and baked products texture.New York: AVI, 1990: 157330.
74. HY Roos. M Karel. Applying state diagrams to food processing and development.
Food Technol 4 3 l2):66, 68-71, 107, 1991.
75.TLJames.NuclearMagneticResonance
in Biochemistry:NewYork:Academic
Press,Inc.,1975.
76. H Pfeifer.Nuclearmagneticresonance andrelaxationofmoleculesadsorbed
011
solids. In: p Diehl. R Ksfeld, eds. NMR: Basic Principles and Progress. New York:
Springer-Verlag, 1 9 7 2 5 - 153.
77. WP Slichter. NMR studies of multiple relaxations in polymers. J Poly111 Sci 14:3348, 1966.
78. HY Roes, M Karel. Water and molecular weight effects on glass transitionsin amorphous carbohydrates and carbohydrate solutions. J Food Sci 56: 1676- 1681, 1991.
79. YH Roes, M Karel, JL Kokin. Glass transitions in low moisture and frozen foods:
effect on shelf life and quality. Food Technol SO( I I ):95- 108. 1996.
Ultrasonics
John Coupland
David Julian McClements
University of Massachusetts, Amherst, Massachusetts
1.
INTRODUCTION
Sound waves are transmitted through materials as perturbations in their physical

structure. Hence, it is often possible to relate the ultrasonic properties of a material to useful information about its macroscopic and microscopic composition
and structure. This chapter introduces the physics of high-frequency sound and
principles of ultrasonic measurementof food properties. Applicationsto real food
materials (solutions, polymers, dispersions,and muscle and plant foods) are then
discussed. Finally, someof the many possible untapped applications of ultrasonic
sensors are introduced.
The absorption or speed of various types of radiation is characteristic of
the properties of the material through which it passes. This is most commonly
exploited with electromagnetic radiationin the well-known formsof spectroscopy
used in the nondestructive evaluation of foods (e.g., infrared, ultraviolet, visible)
( I ) ; however, mechanical waves may also be used. Mechanical spectroscopy is
most widely known at the low frequencies used in small deformation rheological
measurements, but higher frequencies are also valuable, importantly ultrasonics
(-20 kHz to 100 MHz).
Ultrasonic spectroscopy shares two common features with
all spectroscopic
of no value to a working food
techniques. First, the actual measurements are
scientist in their own right. Their practical importance arises from correlations
or processing pabetween the spectroscopic measurement and practical quality
rameters. The relationships are most often empirical but can also be analytical.
217
Coupland and McClements
218
However it is established, the relationship between how a consumer and a spectrometer see a material is likely to be weak. An analytical basis for any technique is therefore preferable because it clearly identifies the relationship between
material properties and instrumental readings. Second, different regions
of the
spectrum are sensitive to different physical and chemical structures present.
In
general, a higher frequency sees faster events, which most typically occur at
smaller scales.
Mechanical waves are a series of mechanical disturbances that propagate
as stresses and strains in the physical bonds of the material. The speed and efficiency of the transmission is sensitive to the nature of the bonds and masses of
the molecules present and therefore to composition. There are two distinct types
of ultrasonic waves; the most commonly used in food nondestructive evaluation
(NDE) arelongitudinalwaves(Fig.
la). In this case,thedeformations of the
material occur in the direction of transmission of the wave. In the second case,
a shearing action, causing
shear waves, the wave passes through the material with
deformations normal to the movementof the wave front (Fig. lb). Combinations
of shearing and longitudinal propagation are also possible. Shear waves are very
strongly attenuated in fluids, and because they cannot propagate far, they are very
rarely used to characterize food materials (typically largely liquid).
It is important to distinguish between the low-powered ultrasound used for
NDE of materials and the high-powered ultrasound used for homogenization,
welding, cell disruption, etc. In sensing applications, the deformations caused by
the passing wave are small-ideally within the elastic limit
of the material and
Direction of Propagation
Fig. 1 Diagrammaticillustration of themodes of vibration in (a) longitudinaland (b)

shear waves.
Ultrasonics
219
hence nondestructive. The large energy levels used in high-powered ultrasound

cause small transient air bubbles to form in the material (i.e., cavitation), which
implode causing large shearing forces and disrupting the material.
Ultrasonic testing reveals certain material properties not readily available
by other techniques (e.g., fast kinetics, microstructure of optically opaque materials-see below). However, their important value in food NDE is that they can
be readily made in optically opaque materials (e.g., meat, milk, chocolate) and
through the walls of pipes, containers, and many packaging materials. The most
significant practical limitation of ultrasound is that it is highly attenuated by gas
cells in a sample. In practice that means it is difficult to transmit high-frequency
ultrasound (> 20 kHz) through many real foods (e.g., most fruit, dough, some
cheese). Additionally, ultrasonic measurements are quick and easy to make, they
can be easily automated for on-line use as part of a process control system, and
they are easily made to a good degree of precision.
An ultrasonic wave passing through a material can be expressed in terms
of its velocity and attenuation. Concisely, this relationshipis given by a complex
wavenumber, k = o / c i a , where c is the ultrasonic velocity, o is the angular
frequency (= 2 x 0 , f is the frequency, i = 4 - I , and a is the attenuation coefficient.* The wavenumberis related to material properties via the following equation ( I ) :
where E is the adiabatic elastic modulus of the material,' which is equivalent to

C,p/C, for a gas or K for a fluid, p is the density, C, and C, are specific heats
at constant pressure and volume, respectively, and K is the bulk modulus. When
a beam of ultrasound passes through a bulk solid, there is some shearing at the
beam edges and K is replaced by K + 4/3G, where G is the shear modulus. This
relationship gives longitudinal ultrasonic velocity measurements some sensitivity
to material shear properties, but as typically K >> G, they are hard to measure.
All the material parameters in this equation are complex, with thereal and imaginary parts containing the storage and loss information of the wave, respectively.
In many cases it is possible to neglect the imaginary component and rewrite Eq.
( I ) in the more widely known form (K is the adiabatic compressibility = K"):
* The attenuatlon coefficient of a material can be expressed as Nepers or decibels per meter, Np.m
or dB.m", respectively. where 1 Np = 8.686 dB.

' This should be distinguishcd from the isothermal elastic modulus normally measured in static loading experiments when heat generated has time to dissipate.
220 McClements
and
Coupland
The compression of fluids by a sound wave causes changes in the physical

alignment of the molecules, and sound energyis lost to heat via conduction from
hot (compressed) to cold (rarefied) areas and by the friction
of one molecule
against another. These mechanisms are knownas thermal and viscous dissipation
losses, respectively, and their contribution to measured attenuation
is given by
classical scattering theory (1):
where a,is the classical attenuation coefficient, y = C,/C,, 6 is the thermal conductivity, and q is the viscosity. In systems whereit is applicable, classical theory
can be usedto measure any of these useful physicochemical properties. However,
often the measured attenuation is higher thanthat predicted classically due to the
scattering of sound by small particles (see Sec. 1II.D) or the presence of certain
chemical equilibria.
Additional energy can be lostif there are chemical equilibria present whose
position is affected by the ultrasonic wave. These additional (nonclassical) losses
can be measured and have been applied to the measurement of fast chemical
a material at equilibkinetics (2). When the compression wave passes through
rium, the change in temperature and pressure displaces the balance of reactants
and products. The equilibrium seeks to reestablish itself in the new conditions,
and its capacity to do so depends on both the speed of the reaction and the frequency of the sound. At low frequencies the temperature-pressure gradients are
so slight that the reaction remains in equilibrium and at high frequencies the
reaction cannot proceed fast enough to respond to the rapid fluctuations. Under
these conditions there is little excess sound absorption, but at intermediate frequencies thereactionposition
is constantlyshiftingandthere
is alargeabsorbance peak and a corresponding relaxation in velocity. By measuring the energy loss (attenuation) as a function of frequency over the relaxation process, the
rate constant, k,, of the reaction is given by:
1
2nfC= - " 2 k , d s
z
where Kc is the equilibrium constant,f, is the center frequency of the relaxation,
T is the relaxation time, and C is the concentration. The rate constant can then be
calculated froma plot of relaxation time against the square root of concentration
if
the equilibrium constant can be calculatedby macroscopic methods. This method
is appropriate for reaction times of the order 10-5-10"" s. Slower reactions can
also be followed if the reactants have a different velocity than the products, although other methods such as optical spectroscopy are generally preferred.
Ultrasonics
221
II. METHODS
A varietyof experimental designs have been developed to measure the ultrasonic
properties of food materials (2,3).
All share the common elements
of an electrical
signal generator, which is used to generate vibrations in an ultrasonic transducer,
and another transducer (or the same one after a time interval) to reconvert the
acoustic energy back to an electrical signal, which is then digitized for analysis.
A. Pitch-and-CatchMethod
Perhaps the simplest and most widely used implementation of these elements is
the pitch-and-catch or sing-around pulsed sound method (3-5). The two
major variations of this device are (a) pulse-echo-the sound is reflected from
a metal plate and detected at the original transducer (Fig. 2b)and (b) through
transmission-the sound from one transducer is detected by a second (Fig. 2a).
If the ultrasonic properties of the material are reasonably frequency independent
(nondispersive), velocity can be calculated from the time taken for the pulse to
travel the known pathlength (from a water calibration) and attenuation from the
(1). If required, the frequency depenlogarithmic decrease in energy with distance
Precise
Pathlength
Imprecise
Pathlength
2:
Fig. 2 Diagram of some typical methods of making ultrasonic measurements: (a) and
(c) are pulse echo methods using a single transducer to produce and detect the acoustic
signal; (b) and (d)are through transmission methods usingone transducer to produce and
another to detect the signal. (a) and (b) are methods using a fluid cell which can be precisely calibrated; (c) and (d)are measurements on irregularly shaped objects whereit may
be impossible to precisely know the pathlength.
222
and
Coupland
McClements
dence can be calculated by using

a finite number of cycles of pure-frequency
A/C voltage to excite the transducer and measuring the propertiesof the singlea
frequencysoundgenerated (i.e., toneburst).Byrepeatingthemethodwith
range of frequencies, it is possible to measure a full spectrum. Alternatively, an
electrical spike can be used to excite a broad-band transducer, which generates
a narrow pulse of sound containing a range of frequency components. Using a
fast Fourier transformation to compare the frequency content of the signal before
and after transmission through the material,
it is possible to measure a region
of the spectrum around the center frequency
of the transducer from a single
pulse (5).
More precise measurements can be made using an interferometer ( 1 ) or a
fixed pathlength resonator (6,7). Both methods set up a standing (continuous)
as a function of pathlength
wave in the sample cell and measure the intensity
(interferometer) or frequency (resonator). The received intensity varies as the
pathlength or frequency is continuously altered, forming nodes and antinodes at
the detector. From the intensity variation the velocity and attenuation may be
calculated.Thesemethodsaretypicallymoreprecisebutslowerthanpulse
methods.
In all cases the parallelism of the cell, the reflectance at all interfaces, and
the energy loss due to beam spreading must be considered. Good temperature
control is also essentialin precise ultrasonic experiments; +O.l"C should be consideredaminimum.Alternatively,thetemperaturecanbemeasuredsimultaneously with the ultrasonic signal using either a thermocouple or a second measurement of a water-filled cell in close thermal contact.It is important to use a fast
enough data capture system (oscilloscopeor analog-to-digital card) to retrieve all
the information in the signal. A good rule of thumbis that the capture rate should
be five times the highest frequency component. Postcapture data processing, such
as averaging and Fourier-domain smoothing, are frequently used to improve the
signal-to-noise ratio.
The methods set out above are very precise but only suitable for liquid
foods or solids with appropriate dimensions. Many foods are solid with irregular
shapes or are too large for easy containment
in a sample cell. In these cases it
is rarelypossible to make veryprecisemeasurements or in some cases even
measure absolute valuesof velocity and attenuation. However, useful information
can often be obtained from the relative position of signal features or low-precision
measurements. Measurements can be made using variations
of the pitch-andcatch/reflectometer method described above; a single transducer is held against
the sample (Fig. 2c) or a pair is clamped around it (Fig. 2d). The pathlength can
be measured usinga micrometer or calipers. Signal qualityis frequently improved
by coating the material under investigation with an ultrasonic coupling fluid (either water or a proprietary gel) to eliminate an air gap that would otherwise
Ultrasonics
223
attenuate the signal. If transmission measurements are impossible, another approach is to measure the reflectance coefficient (proportion of normally incident
energy reflected) at an interface and use it to calculate the ultrasonic properties
of the materials from the following equation:
1 into
where R,? is thereflectioncoefficient of a wave passing from material
material 2 and z is the acoustic impedance of the material (=cp) ( I ) . Reflectance
measurements have also been used to measure the surface smoothness of food
materials (8).
B. Imaging
Most people are familiar with the use of ultrasonic imaging techniques from their
medical applications, for example, in prenatal care. These same principles have
been applied extensively to foods to provide information on the spatial heteroa transducer
geneity of food components. Acoustic images can be generated from
fixed to a robot arm that is placed in a tank filled with a suitable couplant (e.g.,
water) with the material under investigation. Operating
in either pulse-echo mode
or through transmission mode (Fig. 2 ) , the transducer records echoes from the
front surface, back surface, and internal structures in the material from a series
of X-Y positions. Each of the recorded waveforms is known as an A-scan.
A set of A-scans can be used to generate an image by assigning a color to either
the magnitude of the signal at either a fixed time (B-scan) or a selected feature
of the signal (e.g., second echo magnitude, time between successive echoes) (Cscan). The B-scan represents a slice through the material, whilea C-scan is more
useful for identifying the changing properties of a component. Both imaging approaches discard a large quantity of the information in each A-scan and should
be used critically. A diagrammatic illustration of image acquisition is shown in
Fig. 3.
In many cases, most especially large (e.g., whole carcasses and animals)
and water-sensitive materials, it is not appropriate to use the scanning tank approach described above. Medical imaging techniques developed forin vivo measurements on human patients (9) have proved useful in these cases-particularly
for muscle foods.
By using very high-frequency ultrasoundit is possible to achieve resolution
approaching optical microscopy. Acoustic microscopy has been used on occasion
with foods and other soft materials, for example, the detection
of sealworms
and bones in cod fillets (10).
224
A-scans
(time of second
marked echo). Good
data recovery
B-scan
(amplitude in
window)
Some data loss
Fig. 3 Diagram showing how an ultrasonic image is generated using a one-dimensional

image of a model solid with indentations cut in the back wall. The instrument generates
a series of A-scans at different X positions ( I , 2, and 3) then generates an image based
on the signal amplitude at either a set time in the A-scan (B-scan image) or of a selected
feature (C-scan image). Both B- and C-scans are data-reduction methods that may give
misleading results.
111.
APPLICATIONS
A.
Simple Solutions
1. Binary Mixtures
One of the most successful groupof applications of ultrasound in food characterization is the determination of the composition of binary mixtures. The ultrasonic
velocity in ideal mixtures can be calculated as a volume-weighted sum of the
in Eq. (2). Nonideal behavior is an
density and adiabatic compressibility used
indication of association or segregation of components of the mixture and is difficult to predict a priori. So, a more practical approach to concentration determination is to prepare a standard calibration curve and use this for similar unknown
samples. Some typical velocity-concentration curves for common food materials
are shown in Fig. 4. Ultrasonic velocity measurements have been usedto measure
the solids concentration in fruit juices (1 1)and can easily be used to measure the
concentration of most two-component mixtures [e.g.. salt in brine (12), alcoholin
spirits (13), solids in skimmed milk (14)].
225
Ultrasonics
1650
"
1"
5
10
15
20
Concentration (weight%)
25
Fig. 4 Concentration dependenceof velocity for a variety of solutes at 20C;(0)sodium

chloride, (0)glucose, and(W) ethanol. (All data from Ref.73 or measured by the authors.)
By making velocity measurements of concentration as a function of time

and position using an ultrasonic velocity-based imaging device,
it has been possible to measure the diffusion coefficient of sucrose in xanthan solutions (15).
2. Ternary Mixtures
For many simple solutes (e.g., salts and sugars) thereis little temperature dependence in the concentration incrementof velocity, butin other cases (especially fats
and alcohols) the temperature increment is negative while that of water is positive (T < 76C). In the latter cases, at a critical temperature the speed of sound
in the solute is identical to that of the solvent, and velocity is independent of the
solutioncomposition.Thisproperty
is veryuseful in concentrationmeasurements. Consider solutes 1 and 2 , the former sharing a critical point T, with the
solvent; then cTC= f($z) and cT+TC= ($,, @).By developing two calibration curves
at the two measurement temperatures, it is possible to measure the composition
to measure alcohol
of a three-component system. This approach has been used
and solids in wine (13), fat and solidsin milk (14), and fat and proteinin fish (16).
In the absenceof a critical point, orif temperature is not variable, multicomponent
mixtures require additional nonultrasonic measurements for complete characterization; for example, Anton-Paar (Graz, Austria) have developed a nondestructive
method based on simultaneous velocity and density measurements.
B. Lipids
1. Liquid Oils
Food lipids are a mixture of various types of triacyl glycerols along with minor
components including cholesterol, mono- and diacyl glycerols, and vitamins. The
226
velocity of an oil is the volume-weighted sum of the velocities of the component

in quality
fatty acids (17). Oilused for deep-fat frying progressively declines
(18)
throughuse as it partiallyoxidizesandpolymerizes.LaceyandPayne
showed that ultrasonic velocity (at 2.25 MHz) in corn oil increased from 1444.8
of deteriorato 145 1.1 m.s" with frying time and correlates with other measures
tion but was insufficiently sensitive to detect product defects. An empirical relationship has been developed for ultrasonic velocityin oils as a function of refractive index, density. and iodine value (19).
2. MeltingBehavior
The pure chemical components of food oils have a wide range of melting points
in the range commonly encountered during food processing, use, and storage. As
they are a mixture, the combinationof colligative properties and mutual solubility
means the observed melting behavior of food oil occurs over a wide temperature
range (20). The solids content of fatty foods is related to their perceived quality
(e.g., gloss in chocolate, stabilityof emulsions, textureof butter). Ultrasound can
be used to measure the volume fraction of solid fat mixed with liquid oil as the
velocity of sound is much less in liquids than in solids. A typical melting profile
for a sample of chocolate is shown in Fig. 5. The solid fat (SF) content can be
calculated as:
"_
cf ct
1200 LO
10 20 30 40 50 60
Temperature (OC)
Ultrasonics
227
where c is the measured ultrasonic velocity and c, (c,) the velocity in pure solid
fat (liquid oil) extrapolated to the measurement temperature. This equation was
developed from Eq. (2)
by assuming solid fats and liquid oils have similar density
and behave ideally as a mixture (21). Ultrasonic measurements give very similar
data to conventional methods such as NMR and DSC (22). This technique has
been applied to the measurement of solid fat in adipose tissue (21) and oil-inwater emulsions (22).
C. Polymers
Polymers and their aggregates play an important role in determining the stability
and textural characteristics of many foods, and there have been several attempts
of polymer
to use ultrasonic measurements to characterize the bulk properties
networks and the structure of isolated molecules in solution. The attenuation of
sound by a hydrocolbid solution is due to classical, scattering, and relaxational
(fast physicochemical reaction) losses. By measuring the attenuation overa wide
frequency range, it is possible to some extentto separate these effects and ascribe
changes in attenuation to molecular and scattering events. Unfortunately, the relaxations occur over a very wide frequency range and several instruments are
required to capture an entire spectrum. In oneof the most complete studies, Choi
and coworkers (23) measured the spectrum of bovine serum albumin from 0.11600 MHz at pH 1.5-13.2 using five techniques to cover the entire range. They
noted excess absorption at acid and alkali pH due to carboxyl and amino group
proton exchange reactions and structural fluctuationsin the molecule. Other studies using single or narrow frequency ranges for attenuation measurements are
less able to define which molecular events are causing the measured changes but
have had some empirical success.
Audebrand and coworkers (24) studied the gelation of alginate and amylose
by ultrasonic spectroscopy. While velocity was unchanged during gelation, attenuation increased in a manner similar to the real part of complex viscosity (G')
and, in the caseof amylose, turbidity. The time axis of the functions was different
to different molecular profor the three assays consistent with their sensitivity
cesses. Attenuation was shown to become more dependant on gelation at higher
frequencies (100 > 80 > 50 MHz). Small changesin 5 MHz velocity (-2 m.s")
were observed at temperatures close, but not identical, to the gel point of gellan
measured by a mechanical test (25). Measurements of velocity and attenuation
of a-amylase
at lower wavelengths (2 MHz) were used (26) to measure the action
on starch. The attenuation of the material decreased linearly with the number of
bonds broken, and this was ascribed to the (unmeasured) change
in viscosity;
measured velocity was unaffected by enzymatic action.
Coagulation of casein micelles to form a self-supporting network is a crucial stage in cheese manufacture. After a period of reaction, the cut-time, the
coagulum is cut and excess water allowed
to drain out. The cut-point is often
228
defined by the expertise of the cheese maker but can be defined as the time at
which the viscosity of the material sharply increases. Ay and Gunasekaran (27)
showed that the ultrasonic attenuation coefficient
a , at a frequency of 1 MHz,
of milk decreases at a decreasing rate with coagulation and the turning point of
a polynomial equation fitted to the measured data (daldt = 0) provided a similar
time to the accepted rheological method. Velocity showed no significant change
during coagulation, although there wasa very large variation in reported data (of
the orderof ? I O m.s) (28). In another studyof protein aggregation, the attenuation coefficient of solutions of broad bean legumin proteins reached a maximum
at pH values near the isoelectric point
( - 9 , probably because the proteins formed
aggregates that scattered the sound (29). This approach was also used to study
the effects of dextran on limiting the isoelectric precipitation of the same protein
(30).
In summary, it seems that high-frequency attenuation is most sensitive to
the state of food polymers and hencemany of their bulk properties. The velocity
of sound in polymer solutions is largely frequency independent and relatively
insensitive to aggregation.Velocitychangeshavebeenexploited
to a much
greater degree in measurements of individual molecular compressibility via Eq.
(2). The compressibility of a molecule in solution is measured as the change in
solution compressibility on adding one molecule of solute to pure solvent; in
practice this is achievedby extrapolating measurements made in a series of dilute
solutions. [It may also be possible to make measurements at higher concentrations, therefore requiring lower precision, if the scattering of sound is accounted
for (31).] The compressibility of a molecule in aqueous solution depends on (a)
its intrinsic compressibility and (b) the compressibility
of the associated water
molecules and is therefore very sensitiveto the hydration of polymers in solution.
The intrinsic compressibility of the protein (believed to be due to a cavity in
the molecular structure) is less than the surrounding water, while the bound surface water is less compressible than the bulk (32). This model gained support
from a molecular dynamics simulation (33), which further suggested that the
intrinsiccompressibility of the polymer (i.e., the cavity) was identical for
the two globular proteins studied (superoxide dismutase and lysozyme).
If this
observation is generally true for globular proteins, then ultrasonic measurements
can be used to directly measure their hydration state.
The surface hydrationof a protein is largely a measure of surface hydrophobicity; an important parameter governing the functionality of food proteins (34).
It is therefore unsurprising that compressibility correlates with protease susceptiof unfolding (35) and that there is a
bility, foaming capacity, and free energy
measurable change in compressibility on thermal or guanidine hydrochlorideinduced denaturation (32). However, empirical attempts to predict the compressof its constituent amino acids have met with
ibility of a protein from the properties
only limited success (36), suggesting that we are a long way from understanding
mechanistically which properties of a protein are measured by compressibility.
Ultrasonics
229
D. Dispersed Systems
Many foods are dispersed systems, importantly emulsions (e.g., mayonnaise, soft
of colloidal
drinks),foams (e.g., beer,carbonateddrinks),andcombinations
structures (e.g., bread dough, ice cream). When waves pass through the material
inhomogeneities, there is an interaction known as scattering, which is dependant
on the physicochemical properties of the two phases as well as their size, shape,
concentration,spatialdistribution,andthefrequencyoftheultrasoundused.
Some of the ultrasound is directed out of its path and so is not detected, and
some is lost as heat as the scattering is not completely efficient. Consequently,
the measured ultrasonic frequency spectra contain a relaxation, which can be
related back to the physical properties of the material under evaluation.
In certain
cases it is possible to understand the acoustic lossesin terms of various scattering
theories, but if this is not possible, empirical relationships may frequently be
developed.
1. Emulsions
Scattering of sound by emulsion droplets is relatively easy to solve analytically
as the particles are spherical, and their typical size (-pm) is much less than the
wavelength of the ultrasound (-mm), so the long-wavelength approximations to
scattering theory are applicable (37). Under these conditions, sound is scattered
by emulsion droplets by two important mechanisms:
I.
Thermalscattering: The particleandsurroundingmediumarecompressed to different extents by the wave, and the resulting temperature
difference causes a heat flux. Thermally scattered energy radiates in
all directions around the particle (Fig. 6).
2. Viscous scattering: The particle oscillates in the pressure gradient because it has a different density than the continuous phase (Fig.6). This
oscillation leads to the generation of a secondary wave by the particle
that has a cosine dependence on angle.In addition, the particle oscillation is damped by the viscosity of the surrounding liquid and some of
the ultrasound is lost as heat.
Neither mechanism is completely efficient, and there
is significant energy loss.
Using relatively few assumptions, it is possible to develop an analytical expression for scattering from a single particle.
The effect of finite concentrations of
particles can then be accounted for using scattering theory
(38) or a core-shell
a function of particle
model (39) to calculate the bulk ultrasonic properties as
of the component
size distribution and concentration and the physical properties*
phases.
* Viscoslty. density, thermal conductivity, thermal expansion coefficient, specific heat, and ultrasonic
velocity and attenuation.
230
Monopolar scattering
_".
"....___
Dipolar
scattering
Droplet
Fig. 6 Diagram illustrating the scattering of ultrasonic waves by emulsion droplets. The
main mechanisms are droplet pulsation due to differences in thermal properties with the
continuous phase and oscillation due to density differences. These mechanisms scatter
monopolar and dipolar waves respectively.
The bulk properties of common food are well documented in the literature
(40), so, for given ingredients, it is possible to predict the ultrasonic properties
of any size/concentration emulsion. Typical results for a model food emulsion
are shown in Fig. 7.
At all frequencies the velocity and attenuation are dependent
on concentrathe spectra upor down),but over a critical
tion (changing the concentration moves
range of frequencies there is also dependence on particle size. Therefore, using
n
-".
1460
0.03 a,
0
w.
0
,E 1450
0.02
C
0.01
.-0
5
C
0.00
B
a
-7 -6 -5 -4 -3 -2 -1 0 1 2 3
log df
Fig. 7 Velocity and attenuation ofa 10%corn oil in water emulsion ( his the wavelength
of the sound, other symbols are defined in the text). (Calculated as described in Ref. 38
using data from Ref. 40.)
Ultrasonics
231
high- or low-frequency measurementsit is possible to measure droplet concentration and use this value to measure size from measurements in the central region.
It would of course be preferable to record the entire spectrum and solve for size
and concentration simultaneously. Particle size measurements using either velocity or attenuation agree with laser diffraction scattering measurements
in pmin concentratedemulsions(volumefraction,
Q < 0.5)
sizedfoodemulsions
(41,42). This is a particularly important application of ultrasonic NDE, as the
information obtained cannot be readily measured by other methods. Commercial
instruments based on these principles are available from various suppliers.
An interesting extension of this method arose from one of its limitations.
When the scattered waves from one particle interact with those from another
(multiple scattering) less energy is lost. This occurs when the average particle
separation decreases to a critical level, either at high concentrations when the
measured attenuation increases less rapidly than the simple theory would predict
or in flocculated emulsions. Detection of flocculation in emulsions is particularly
of physical deteriorationof a product.
important, as it is frequently the initial stage
McClements (43) was able to detect flocculated emulsions with this method before they began to visibly cream.
When the particles are charged, there is additional attenuation due to ionic
friction between the moving particle and its counterions, which generates an
A full soluA/C voltage that can be measured alongside the acoustic attenuation.
tion of the viscous, thermal and electroacoustic scattering losses allows calculation of particle size and surface charge (c-potential) (44). This method showed
good agreement with previously published values for casein micelles
in skim
milk (45).
Ultrasonic imaging has been widely used to study the creaming of food
emulsions under gravity. In its simplest form this method merely relates the vea known pathlength as a function of time
locity for the sound to pass through
and position to the volume fraction (46), but by measuring the scattering effects
it should also be possible to determine size separation under gravity (47).
2. Foams
The concentration, size, and growth of air cells
in bread, fruit, dairy products,
beers, and wines are vital to the perceived quality of these products. Ultrasound
is very sensitive to dispersed gases and would seem an ideal investigative tool,
but in practice the attenuation due to the resonant scattering of the bubbles is so
strong that transmission measurements are not possible at high frequencies
(>O. 1
MHz). Measurement of the surface reflection coefficient is a more practical apa
proach to capturing the important frequency dependence of scattering from
bubbly liquid, and the potential of this method has been demonstrated for some
whipped food materials (48).
232
E. Muscle Foods
Ultrasound has been used for a number of years to measure the fat content of
various live animals and carcasses. Such whole animal studies are beyond the
scope of this work, but the principle
of the measurement is similar to the unknown
pathlength pulse-echo method described above (Fig. 2d). A transducer is
held
against the back of the animal and a pulse of sound fired through the surface
layer and an echo recorded from the fat/muscle interface. The fat thickness corre(49-51).
lates with overall fat content and other carcass and meat properties
Alternatively, pattern recognition techniques developed for medical imaging devices may be used (52,53).
Another imaging method that has seen some success
in imaging muscle
foods is elastography(54). In this method, an A-scan is recorded before and after
the material is slightly compressed
by the transducer. Pressing the transducer into
the material causes the material to be deformed, the softer materials more than
the harder, and the relative movements of the peaks can be tracked by crosscorrelation techniques. In this method isitpossible to get imaging across a plane
in the material based on Youngs modulus. Ophir and coworkers(54) have used
this approach to distinguish between fibrous muscle and perimysial tissue and to
visualize a healed traumatic injury in beef samples. By calibrating the analysis
with material of known properties, it is possible to make absolute measurements
of Youngs modulus. A typical elastography image of a meat sample is shown
in Fig. 8, in which the light bands represent bands of collagen and the dark areas
fibrous muscle.
Fig. 8 Elastographic image of beef muscle; differentiation is based on the elastic modulus of the material. The white areas represent connective tissue and the dark myofibrillar
muscle. (Image kindly donated by Rhonda Miller, Texas A&M University.)
Ultrasonics
233
Muscle tissue can be considered a combination of a protein solution and

fat, and the speed of sound usually lies somewhere between the values for oil
and water (- 1400- 1500 ms). The fat concentration can therefore be estimated
fairly accurately for meat ( 5 5 ) and fish (56) using the methods set out above for
it is possible to measure
simple solutions. Using the critical-point approach,
fat, protein, and moisture in fatty fish tissue (1 6). Attenuation is a less reliable
measure of muscle composition, as it is very sensitive to microstructure. For
example, there isa very large increase in the attenuation coefficient of fish tissue
after a freeze-thaw cycle (D J. McClements, unpublished data), probably because
small air bubbles forced outof solution by the freezing processdo not completely
redissolve.
This approachis limited, as it is difficult to make accurate velocity measurements of real meat cuts in a processing environment. One solution to this problem
is to consider the frequency-domain energy distribution of a transmitted broadband pulse of ultrasound. Despite requiring less information about the measurea strong relationship (r = 0.89)
ment system (pathlength not considered), there is
between the number of frequency-domain peaks and fat content of beef muscle
(57); weaker correlations were observed for other signal features. The same frequency-domain approach showed some correlation with the sensory perception
as a predictive tool
of juiciness, flavor, and texture but were too weak for use
(58).
F. Plant Foods
The ripening and deterioration of vegetable products is associated with changes
in chemical composition and mechanical properties that might reasonably be expected to change the ultrasonic properties of the material. However, the correlations between ultrasonic and quality parameters are often weak for fruits and
vegetables because the theoretical link between acoustic propagation and strucof
ture is poorly understood. The acoustic properties are an unknown function
vegetable material properties including size concentration and distribution of air
cells, the cytoplasmic composition, the mechanical properties
of the cell walls,
and the intercellular bonds, while the perceived quality (e.g.,flavor, crunchiness)
is probably a very different function of physicochemical structure. Methods that
more closely mimic the consumers use of a food (e.g., compressional tests, GC
analysis of volatiles) are likely to correlate better with perceived quality, but
because these are inevitably destructive, acoustic methods have been frequently
considered. A good example of the limitation of acoustic measurements was seen
in recent (59) measurementsof the velocity, attenuation(37 kHz), and other properties of carrots cooked for different times (0-15 rnin). Measured velocity decreased linearly with strain at failure, Youngs modulus, solids content, and density, butthecorrelationswereweak
(r = -0.69,-0.62,-0.46,and-0.29
234
respectively), while attenuation showed still weaker positive and negative correlations with the same parameters.
Practically, vegetables are difficult to measure because not only are they
irregular in shape and variable in composition, but they also frequently contain
intercellular air cells, which scatter ultrasound and cause unmeasurably high attenuation at high (->20 kHz) frequencies (60). Adequate transmission measurements are possibleat lower frequencies, butthis is not ideal as there is less spatial
resolution and beam spreading and wave-guide effects can reduce the precision.
in the
Despite these limitations, many researchers have reported some success
ultrasonic NDE of plant foods. Cheng and Haugh (61) identified whole potatoes
with hollow heartas they absorbed more low frequency (0-75 kHz) sound energy
than the healthy samples. The very-low frequency sonic resonance (<1 kHz) of
whole apples decreased monotonically during the accelerated ripening process
in a similar manner to texture measured bya puncture test (62). Self and coworkers used low-frequency ultrasonicsto follow ripening in banana (60) and avocado
(63) and to determine the elastic modulus of apple tissue (64).
G. Miscellaneous
One of the very few concerted uses
of shear wave ultrasound in food systems
was made by Lee and coworkers (65), who made measurements on cheese and
dough (0.5-1 MHz) and comparedthe results with conventional oscillatory rheological measurements. Both devices showed similar trends in complex rheological properties with composition, but because the devices operated over different
frequency ranges, quantitative comparisons could not be made. The success
of
thisstudy is surprising, because ultrasonic waves
of this frequency would be
expected to be very highly attenuated in a gas-containing mixture such as dough
and shear waves would not be expected to propagate in fluid media.
The time of flight for a 1.25 MHz ultrasonic pulse across a pipe was used
to detect fouling at the pipe surface (66).
The sensor could measurefilms between
0.5 and 6 mm thick, the lower limit being set by the wavelength
of the ultrasound
and the upper limit by the acoustic impedance
of the materials. When implemented in a pilot-scale UHT milk plant, results were unaffected by flow rate (0I O L/min).
The crispness of several types of cookies and crackers as measured by a
sensory panel and fracture testing correlated with ultrasonic velocity (67). Ultrasonic absorption has been used to measure the texture of wafer sheets (68).
There have been some attempts at using ultrasonic imaging to detect the
growth of spoilage organisms in packaged milk (69) and other liquid foods (70).
Good images could be recorded through plastic and metal packaging materials
(paper contains air cells that strongly attenuate the sound) (70), and the speckle
density could be correlated
to conventional plate counts. However, only very
Ultrasonics
235
large numbers of bacteria could be detected by this method, and a preincubation

of many of
period was required. It was suggested that the proteolytic activity
the bacteria caused the changes detected
by this method, which may limit any
application.
IV. FUTURETRENDS
A survey of the literature in this field reveals more than 40 years of ultrasonic
(71,72). This
methods of foodcharacterizationbasedonultrasonicprinciples
concerted effort has produced remarkably few practical applications, most importantly on-line concentration meters, carcass evaluation tools, and, most recently,
emulsion particle-sizing instruments. This developmental effort must be better
focused to meet actual rather than perceived needsin the food industry. In developing a method it is important to remember the strengths and weaknesses of
a given technique as summarized for ultrasonics in this work. It is also worth
remembering that a new method will not be adopted based on its novelty alone;
it must provide information that is not readily attainable by existing technology
or be in some way more practical (e.g., cheaper, easier
to use). Additionally, food
a measurement device must maintain
is an intrinsically variable material, and
sensitivity to the important variable (e.g., crunchiness in apples) while ignoring
changes in other variables (e.g., variety, size, insect damage). Food-processing
plants are rarely static, single-product operations, and the sensor must be flexible
enough to account for this. For example, a juice-bottling plant may switch bea day, and
tween several product types over different lines over the course of
any analytical controls must be equally easy to recalibrate or their potential for
improved quality and production times will be
lost in the inconvenience of operation.Bearingthesepoints
in mind,wewouldliketosuggestsomeareaswe
believeultrasound-basedinstrumentationcouldbeused
in thenondestructive
evaluation of foods:
Polymer characterization:Very high precision resonator cells and digital

signal generators are now becoming available at moderate cost, and their
of their hydration and
application to food materials will reveal details
small molecule binding that cannot be readily obtained by other methods.
Guided wavetechnology: An ultrasonic wave can be angled using
a wedge
to trap most of the energy within a container or pipe wall. By adjusting
the angleof incidence and the frequencyof the wave, element movement
is excited along the length of the pipe (in phase vibrations) and at right
angles toit (out of plane vibrations). Variationof the proportions of these
vibrations changes the sensitivity of the technique to pipe and contents
properties, respectively. Proper application of this method can be used
236
Coupland and McClernents
to probe many tens of meters of pipe from a single measurement and

can be used to detect corrosion, fouling, weld defects, and properties of
the material contained (J. Rose, unpublished).
Particlecharacterization: Ultrasonicspectroscopyhasbecomeanaccepted technique for the characterization of model and
real emulsions.
However, many of the important functional propertiesof food emulsions
occur when they stop acting as isolated liquid spheres and instead aggregate,gel,orcrystallize.Newdevelopments
in scatteringtheorywill
uniquely allow ultrasonic sensors to investigate these events.
Surface acoustic waves:Using printed circuit technology, it is possible to
A/C
makeminiatureultrasonictransducersthat,whenexcitedbyan
pulse, transmit sound waves across the surfaceof the imprinted material.
in pulseThe waves can be detected by a similar transducer (or used
echo mode) and the speed of the wave measured. The speedof the wave
is proportional to the distance between the transducers and is therefore
sensitive to temperature (via the thermal expansion coefficient) and shear
stress on the material. Modifications of this device are also sensitive to
the chemical environment into which the sensor is placed and can be
of trace chemicals. The
used to detect pH, composition, and the presence
great advantage of these devices is that they are small (< 1 n m ) , cheap
to wireless interrogation. If a
enough to be disposable, and amenable
chip was attached to the packaging of an individual food item, it could
be used to sound an alarm when decay reaches a specified level
or in
conjunction with an intelligent oven to determine if the food is cooked
(V. Varadan, unpublished).
Advanced imaging technologies: Medical imaging ultrasound is an established diagnostic tool that could be easily
used to detect defects and
contamination in food materials. Typical food-processing operations require much faster throughput thanis (thankfully) acceptable in a doctors
office, so the large amount of information obtained must be reduced to
a binary decision variable by computer-based data analysissoftware similar to that used in other machine vision applications. On-line imaging
could be used to detect glass in opaque foods, insect contamination i n
meat products, and mechanical damage to packaging materials.
Shear wave studies: Shear wave velocity is much more sensitive t o the
perceived texture of food than the more commonly applied transverse
waves and is limited by difficulties making adequate measurements in
fluids. Development of an accurate measurement method would effectively provide a noninvasive, on-line viscometer.
I n this chapter, useful information about foodstuffs that can be extracted

from their interactions with high-frequency sound has been presented. New devel-
Ultrasonics
237
opments in theory, instrumentation, and application are constantly emerging

in
the physical sciences literature, and
it is the application of these fundamental
ideas to food materials that will lead
to the improvements in food quality and
safety that would result from good nondestructive evaluation.
GLOSSARY
Attenuation coefficient (a):The loss of acoustic energy per unit distance.
Couplant: Ultrasound is strongly attenuated in air and a liquid couplant (often
water of a proprietary gel) is needed to facilitate the transmission between
the transducer and sample.
Delay line: Apiece of material(frequentlyPlexiglasorquartz)designed
to
separate the large signal from the transducers
initial ring from later measurable echoes.
Dispersed system: A material that is inhomogeneous supramolecular scale but
homogeneous in bulk. Includes emulsions and foams.
Elastography: An imaging system based on the change in detected signal on
bulk deformation.
Emulsion: Aliquid-in-liquiddispersedsystem.
Foam: Agas-in-liquiddispersedsystem.
Frequency: The reciprocal of thetimerequiredforawave
to complete on
cycle. The frequencyof ultrasonic waves usedin food analysisis commonly
in the 10hs" region.
Imaging: Use of the spatial dependence of an ultrasonic signal to develop an
image of internal structure.
Impedance: A measure of the capacity of a material to transmit sound. Commonly measured as the product of density and ultrasonic velocity.
Resonator: An ultrasonic measurement system where the properties of interest
are measured from the constructive and destructive interference
of a continuous wave with itself.
Scattering: The process by which the direction
of transmission of ultrasonic
energy is changed by interaction with microscopic inhomogeneities.
Transducer: Mechanical device to convert electrical to mechanical (ultrasonic)
energy or vice versa.
Ultrasound: Mechanical vibrations similar to sound but at frequencies far beyond the range o f human hearing (>20 kHz)
Velocity (c): The speed of sound in the direction of transmission or detection.
Wavelength: The distance between points of equal amplitude on a wave. Ultrasonic waves used in food characterization typically have a wavelength of
the order of a millimeter.
238
REFERENCES
I . J Blitz. Ultrasonics: Methods and Applications. London: Butterworths, 1971.
2. MJW Povey, DJ McClements. Ultrasonics in food engineering. Part 1: introduction
and experimental methods. J Food Eng 8:217-245, 1988.
3. EPPapadakis.TheMeasurementofUltrasonicVelocity.In:
RN Thurston,AD
Pierce, eds. Ultrasonic Measurement Methods. Vol.
XIX Physical Acoustics. San
Diego: Academic Press, 1990:s 1- 106.
4. EP Papadakis. The Measurement of Ultrasonic Attenuation. In: RN Thurston, AD
Pierce, eds. Ultrasonic Measurement Methods. Vol.
XIX Physical Acoustics. San
Diego:AcademicPress1990:107-155.
5. DJ McClements, P Fairley. Frequency scanning ultrasonic pulse echo reflectometer.
Ultrasonics30:403-405,1991.
6. F Eggers, T Funck. Ultrasonic measurements with milliliter liquid samples in the
0.5-100 MHz range. Rev Sci Instrum 44:969-977, 1973.
7. AP Sarvazyan. Development of methods of precise ultrasonic measurements
in small
volumes of liquids. Ultrasonics 20: 15 1 - 154, 1982.
8. N Sarkar, RR Wolfe. Potential of ultrasonic measurements in food quality evaluation. Trans ASAE 26:624-629, 1983.
9. KK Shung. Basic principles of ultrasound tissue characterization. In: SE Freeman,
ed. Noninvasive Techniquesin Biology and Medicine. San Francisco: San Francisco
Press Inc, 1990:205-217.
10. H Hafsteinsson, SSH Rizvi. A review of the sealworm problem: biology, implications and solutions. J Food Prot 50:70-84, 1987.
1 1 . NI Contreras, P Fairley,DJ McClements, MJW Povey. Analysis of thesugar content
of fruit juices and drinks using ultrasonic velocity measurements.
Int J Food Sci
Techno1 27515-529, 1992.
12. BB Owen, PL Kronick. Standard partial molal compressibilities by ultrasonics.
11.
Sodium and potassium chlorides and promides from 0-30C. J Phys Chem 65:8487, 1961.
13. WC Winder, DJ Aulik, AC Rice. An ultrasonic method for direct and simultaneous
determination of alcohol and extract content of wines. Am J Enol Vint 2 1 :1 - I 1.
1970.
14. JW Fitzgerald, GR Ringo, WC Winder. An ultrasonic method for measurement of
solids non-fat and butterfat in liquid milk. J Dairy Sci 44:1165, 1961.
15. TK Basaran, JN Coupland, DJ McClements. Monitoring molecular diffusion of sucrose in xanthan solutions using ultrasonic velocity measurements. J Food Sci (submitted).
16. R Ghaedian, JN Coupland, EA Decker, DJ McClements. Ultrasonic determination
of fish composition. J Food Sci (in press).
17. DJ McClements. MJW Povey. Ultrasonic velocity measurements in some liquid triglycerides and vegetable oils. J Am Oil Chem SOC 65: 1787-1789, 1988.
18. RE Lacey,FA Payne. Ultrasonic velocityin used corn oil as a measure of oil quality.
Trans ASAE 37:1583-1589, 1994.
19. TH Gouw. JC Vlugter. Physical properties of triglycerides 111: ultrasonic sound velocity. Fette Siefen Anstrichmittel 3:159-164, 1967.
Ultrasonics
239
cc Casmir,
20. PJ Lawler, PS Dimick. Crystallization and polymorphism of fats. In:
DB Min,eds.FoodLipids:Chemistry,Nutrition,andBiotechnology.NewYork:
Marcel Dekker, 1998: 229-250.
21. CA Miles, GAJ Fursey, RCD Jones. Ultrasonic estimation of solid/liquid ratios in
fats, oils, and adipose tissue. J Sci Food Agric 36:218-228, 1985.
22. DJ McClements, MJW Povey. Comparison of pulsed NMR and ultrasonic velocity
measurements for determining solid fat contents. Int
J Food Sci Technol 23:159170,1988.
23. PK Choi, JR Bae, K Takagi. Ultrasonic spectroscopy in bovine serum albuminsohtions. J Acoust Soc Am 87:874-881, 1997.
JR Emery. Investigationof gelation phenom24. M. Audebrand, JL Doublier, D Durand,
ena of some polysaccharides by ultrasonic spectroscopy. Food Hydrocoll 9:195203,1995.
25. Y Tanaka, M Sakurai, K Nakamura. Ultrasonic velocities in aqueous gelIan S O ~ U tions.FoodHydrocoll7:407-415,1993.
26. MJW Povey, AJ Rosenthal. Technical note: ultrasonic detection of the degradation
of starch by a-amylose. J Food Technol 19: 1 15-1 19, 1984.
27. C Ay, S Gunasekaran. Ultrasonic attention measurements for estimating milk clotting time. Trans ASAE 37:857-862, 1994.
28. S Gunasekaran, C Ay. Milk coagulation cut-time determination using ultrasonics.
J FoodProcEng19:63-73,1996.
29. G Pavlovskaya, DJ McClements, MJW Povey. Ultrasonic investigation of aqueous
solutions of a globular protein. Food Hydrocoll 6:253-262, 1992.
30. G Pavlovskaya, DJ McClements, MJW Povey. Preliminary study of the influence
of dextran on the precipitation of leguminin from aqueous salt solutions. Int J Food
Sci Technol 27:629-635, 1992.
31. VJ Pinfield, MJW Povey. Thermal scattering must be accounted for in the determination of adiabatic compressibility. J Phys Chem B 101: 1 1 IO- I 1 12, 1997.
32. Y Tamura, K gekko. Compactness of thermally and chemically denatured ribonuclease A as revealed by volume and compressibility. Biochemistry 34: 1878- 1884,
1995.
in solvated
33. E Paci,MMarchi.Intrinsiccompressibilityandvolumecompression
proteins by moleculardynamicssimulationathighpressure.ProcNatlAcadSci
USA 9311 1609-11614, 1996.
34. S Nakai. Structure-function relationships in food proteins with an emphasis on the
importance of protein hydrophobicity. J Agric Food Chem 31:676-683, 1983.
35. K Gekko, K Yamagami. Flexibility of food proteinsas revealed by compressibility.
J Agric Food Chem 3957-62, 1991.
36 MM Gromiha, PK Ponnuswamy. Relationship between amino acid properties and
protein compressibility. J Theor Biol 165:87-100, 1993.
37. DJ McClements, JN Coupland. Theory of droplet size distribution measurements in
emulsions using ultrasonic spectroscopy. Colloids Surfaces A 1 17: I61- 170, 1996.
in emulsions.
38. DJ McClements. Principles of ultrasonic droplet size determination
Langmuir123454-3461,1996.
39. Y Fukumoto, T Izuyama. Thermal attenuation and dispersion of soundin a periodic
emulsion. Phys Rev A 46:4905-4921, 1992.
240
40. JN Coupland, DJMcClements.Physicalpropertiesofliquidedibleoils.

J AmOil
Chem Soc 74:1559-1564, 1997.
41. JN Coupland, DJ McClements. Droplet size determination in food emulsions: comparison of ultrasonic and light scattering methods. J Food Eng (in press).
42. DJ McClements, MJW Povey, M Jury, E Betsanis. Ultrasonic characterization of
a
food emulsion. Ultrasonics 28:266-272, 1990.
43. DJMcClements.Ultrasonicdeterminationofdepletionflocculationinoil-in-water
emulsions containing a non-ionic surfactant. Colloids Surfaces A 9095-35, 1994.
44. RW OBrien, DW Cannon, WN Rowlands. Electroacoustic determination of particle
size and zeta potential. J Colloid Int Sci 173:406-418, 1995.
45.TWade,JKBeattie,
WN Rowlands, MA Augustin.Electroacousticdetermination
of size and zeta potential of casein micelles in skim milk. J Dairy Res 63:387-404.
1996.
46. PA Gunning, DJ Hibberd, AMHowe, MM Robbins.Useofultrasoundtomonitor
gravitational separation in emulsions. J Soc Dairy Techno1 42:70-77, 1989.
47. VJ Pinfield, E Dickinson, MJW Povey. Modeling of concentration profiles and ultrasonic velocity profiles in a creaming emulsion: importance of scattering effects. J
Colloid Int Sci 166363-374, 1994.
48. PF Fairley, DJMcClements,MJWPovey.Ultrasoniccharacterizationofsome
aerated foodstuffs. Proc Inst Acoust 13%-70, 1991.
49.JCAlliston,
AJ Kempster, MGOwen,MEllis.Anevaluationofthreeultrasonic
machines for predicting the body composition of live pigs of the same breed, sex,
and live weight. Anim Prod 35:165, 1982.
SO. SD Chen. CY Huang, HJ Chen. Prediction of carcass quality in live pigs by ultrasonic measuring of backfat thickness and loin eye area.
11. TaiwanSugar41:24,
1994.
5 I . SDChen,CYHuang,
HJ Chen.Predictionofcarcassqualityinlivepigs
by
ultrasonic measuring of backfat thickness and loin eye area.1. Taiwan Sugar 40:26.
1993.
52.RKolb, G Nilter.Digitizedultrasonicimagesonlivepigsforpredictionofmeat
proportion in the belly. Zuchtungskunde 65:297, 1993.
53. JR Brethrour.Estimatingmarblingscoreinliveanimalsusingpatternrecognition
and neural network procedures. J Anim Sci 72: 1425- 1432, 1994.
54. J Ophir, RK Miller, H Ponnekanti, I Cespedes. AD Whittaker. Elastography of beef
muscle. Meat Sci 36:239-250, 1994.
55. B Park,ADWhittaker, RK Miller, DS Hale.Predictingintramuscularfat
in beef
longissimus muscle from speed of sound. J Anim Sci 72: 109-1 16, 1994.
56. R Ghaedian. EA Decker. DJ McClcments. Use of ultrasound t o determine cod fillet
composition. J Food Sci 62:500, 1997.
57. B Park, AD, Whittaker,
RK Miller,DEBray.Measuringintramuscularfat
in beef
with ultrasonic frequency analysis. J Anim Sci 72:l 17-125. 1994.
58. B Park,ADWhittaker, RK Miller, DS Hale.Ultrasonicspectralanalysisforbeef
sensory attributes. J Food Sci 59:697-701. 724, 1994.
59. M Nielsen, HJ Martens.Lowfrequencyultrasonicsfortexturemeasurcments
in
cooked carrots (Dalruts c o w f a L.). J Food SCI62: I 167- 1 170, 1 175. 1997.
60. GK Self. MJW Povey, H Wainright. What do ultrasound measurements
in fruit and
Ultrasonics
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
241
vegetables tell you?In: MJW Povey, DJ McClements. eds. Developments in Acoustics and Ultrasonics. Bristol: Institute of Physics Publishing, 1992: 129-164.
Y Cheng, CG Haugh. Detecting hollow heart in potatoes using ultrasound. Trans
ASAE 37:217-222, 1994.
LA Liljedahl. JA Abbott. Changes in sonic resonance of delicious and golden delicious apples undergoing accelerated ripening. Trans ASAE 37:907-912, 1994.
GK Self, E Ordozgotti, MJW Povey, H Wainwright. Ultrasonic evaluation of ripening avocado flesh. Postharvest Biol Technol 4: 11 1-1 16. 1994.
GK Self, G Chan, MJW Povey, H Wainwright. Ultrasonic nondestructive evaluation
of the maturity and qualityof fresh fruit and vegetables. In: Ultrasonic International
'91 ConferenceProceedings.Oxford:Butterworth-Heinemann,1991.
HO Lee, H Luan,DG Daut. Use ofan ultrasonic technique to evaluate the rheological
properties of cheese and dough. J Food Eng 16:27-150, 1992.
PM Withers. Ultrasonic, acoustic and optical techniques for the non-invasive detection of fouling in food processing equipment. Trends Food Sci Technol 7:293-298,
1996.
MJW Povey, CA Harden. An application of the ultrasonic pulse echo technique to
the measurement of crispness of biscuits. J Food Technol 16: 167- 172, 198 1.
G Joudiekene. A Petranskas, 0 Reidl, E Mohr. Possibilities of application of ultrasonics to determining the texture of wafer sheets. Ernahrung 14:507-51 I , 1990.
R Ahvenainen, G Wirtanen, M Manninen. Ultrasound imaging-a non-destructive
method for monitoring the microbiological quality of aseptically packed milk products. Lebenstn Wiss Technol 22:382-386, 1989.
R Ahvenainen. G Wirtanen, M Manninen. Ultrasound penetration through different
packagingmaterials-anon-destructivemethodforqualitycontrolofpackaged
UHT milk. Lebensm Wiss Technol 22:268-272, 1989.
DJ McClements. Ultrasonic characterization of foods and drinks: principles, methods, and applications. Crit Rev Food Sci Nutr 37: 1-46, 1997.
MJW Povey. Rapid determination of food material properties. In: MJW Povey, TJ
Mason, eds. Ultrasoundin Food Processing. London: Blackie Academic and Professional, 1998: 30-65.
KH Hellwege. Molecular acoustics. In: KH Hellwege, AM Hellwege, e&. Land&Bornstein, New Series. Berlin: Springer-Verlag, 1967.
7
Firmness-Measurement Methods
Yen-Con Hung, Stan Prussia, and Gabriel 0.1. Ezeike
The University of Georgia, Griffin, Georgia
1.
INTRODUCTION
A.
DefiningTexture
Texture is a basic physical property of foods. Consumers evaluate food texture

as the sum total of kinesthetic sensations derived from eating a food. Perception
of texture during eating encompasses the mouthfeel, the masticatory properties,
the residual properties, and the sound (1). Other definitions of texture include (a)
the physical properties of foodstuffs related to mouthfeel or eating quality (2),
(b) one of the three primary sensory properties of foods that relate entirely (or
in addition to the other primary properties) to the sense of touch or feel (3), and
(c) overall physical properties perceivedby the eyes, fingers, or the mouth during
mastication (4). The primary parameters describing texture have been reported
and adhesiveby Szczesniak (5) as hardness, cohesiveness, viscosity, springiness,
ness.
B.
Importance of TexturelFirmnessEvaluation
Food texture is understood to be the sensory manifestation of kinesthetic properties (6) and depends on particle size, shape, structure, and rheological parameters
( 5 ) . The quality and acceptability of any food are influenced considerably by
textural and structural properties normally associated with it. In relation to food
quality, the importance of texture can be best demonstrated by food product advertisements. Some words that have been used to describe desirable textural quality include crisp, crunchy, tender, juicy, creamy, firm, spongy, and smooth. Conversely, some descriptors that have been usedto characterize undesirable textural
quality include soggy, sticky, gummy, soft, fibrous, tough. and dry.
243
Hung et al.
244
Texture is also one of the key factors in determining harvest date and can
of foods (7). For example, Washalso be used for determining the market grades
(8) requirethatRedDeliciousapplesatthetime
of
ingtonStateStandards
shipping have firmness values
of at least 53 N (12 Ibf) as determined by the
Magness-Taylor (MT) puncture test. Texture has also been used as an indicator
to crackers) also
for freshness of seafood (9). Processed foods (from hotdogs
need to maintain certain well-defined textural quality to meet consumer expectations. In manufactured foods, texture has also been used as an indicator for the
consistency of products.
as measuring
Bourne (10) summarized texture-measurement instruments
either force, distance, time, energy, multiple units, or chemical composition. An
overview was also provided by Giese (1 1) on different food texture-measuring
methods. However, there are more textural parameters than can possibly be covered in detail in one chapter. Firmness is a popular term for hardness ( 5 ) and has
been described as the force necessary to attain a given deformation. Firmness
has also been used as an indication of quality as well as a parameter for sorting
operations. For the remainder of this chapter, the discussion is focused on firmness.
C.Overview
of Firmness-EvaluationMethods
There is a recognized need for development of mechanical devices to separate

(1 2 ) grouped firmness-sensing methods
objects based on firmness. Chen and Sun
as (a) destructive firmness tests, (b) nondestructive methods, and (c) nondestructive measurement of secondary properties. According to Voisey and Larmond
(13), the validity of an instrumental measurement of texture should be based on
it is important to examhow well it predicts sensory perception of the food. Thus,
ine how firmness determined instrumentally (destructive and nondestructive) reon the product
lates to sensory perception of food texture. Firmness also depends
being evaluated. For example, penetrometer readings showedlittle difference between the firmness of apple and muskmelon, even though substantial differences
existed in sensory perceptions. Harker et al. (14) found that the relationship between instrumental and sensory perception of firmness was curvilinear and concluded that further work is needed to fully understand the fundamental psychophysical basis of human perception of texture. They also suggested that sensory
perception of texture is more sensitive than instrumental measurements when soft
fruit are being examined.
The measurement of food texture falls broadly into two categories: instrumental and sensory methods (Fig. 1). The specific principles employed in many
of the instrumental methods are nearly as varied as the products. The methods
can be broadly divided into destructive and nondestructive methods, and the principles range from mechanical to acoustic, resonance, laser air-puff, nuclear mag-
245
Sensory
(Subjective, methodolog ,
psychology, physlologyr
Instrumental
(Objective)
Destructive
(Llmited sample size)
Nondestructive
(On-line, more samples)
Acoustic
1 1 Resonance I I NMR 1 1 Other I

Sensor
Fusion
Human
I Judgment I
Fig. 1 Classification of texture-assessmentmethods.
netic resonance (NMR), etc.Dull ( I 5) defined the term nondestructive to indicate that the sample (of any physical form) can be analyzed to obtain meaningful
of
data by some means in such a way that the physical and chemical attributes
the sample are not altered. This means that the sample, after testing, can be subjected to subsequent uses or activities such as food manufacture, handling, storage, or direct consumption without loss.
After considering the various approaches used for categorizing firmness,
we have decided on two broad categories: methods having mechanical contact
and those having no mechanical contact with the food.
II. MECHANICALCONTACTWITHPRODUCT
A.
DestructiveFirmnessTests
The standard measure of food firmness is conducted with a penetrometer developed by Magness and Taylor (16). It is a mechanically driven probe used to
penetrate the food products with the maximum required penetration force giving
an indication of firmness. In addition to the Magness-Taylor penetrometer, Studman and Yuwana ( 1 7) presented a new firmness-measurement method based on
the moment necessary to rotate a blade attached to a spindle after it has pushed
246
Hung et al.
into the fruit. After the blade has been pushed into the fruit to the desired depth,
the fruit is slowly rotated manually about the axis of the spindle. When the fruit
tissue fails, the maximum angle of rotation is read and the value is used to determine firmness. The resistance to the moment is assumed to be due entirely to
the crushing of the cells by compression and shear. This device requires less than
IS seconds to complete and record each measurement, is able to take measurements at specified depths, and is able to sense a wide range of fruit firmness by
using different sizes of blades for firm (smaller blade) and soft (larger blade)
fruits. There is no need to clean the device between tests, and no need to remove
the fruits skin.
Another modified force-deformation method was investigated by Murillo
et al. ( I 8) using peaches. A cylindrical plug measuring 1.5-2.0 cm long was cut
from each peach 2-3 mm below the skin surface on the blush side using a 2-3
cm (internal diameter) cork borer. Using the Instron Universal Testing machine
with a load cell of 50 kg at cross head speed of 20 mm/min. the specimens were
cut axially through the plug with a 1.36 mm diameter wire probe, and a steadystate force was noted to represent firmness.
The Magness-Taylor methodhas a key advantage of many years of history
that have led to its use as a standard against which other methods are compared
in order to determinetheiracceptability.However,the
main disadvantage of
methods based on this principle is that the probe test is destructive to the product
being measured and only provides test results from a sample that
is used to represent the lot. Therefore, this method cannot beused to measure firmness of every
item going through a packing or processing line.
B. NondestructiveForceMethods
Destructive tests for firmness continue mostly because suitable sensors are not
available for nondestructively measuring firmness
of all items in a lot. Consequently. considerable effort has been made for finding a nondestructive method.
The specific principles employed in many of the instrumental methods are nearly
as varied as the products.
1. Based on Human Touch
As indicated before. manual separation/sortingis the only practical method available now for firmness sorting of foods. Hung et al. (19) used 100 untrained consumers to rank five peaches with different firmnesson a hedonic scale from most
firm, very firm, firm, less firm, to least firm by touching the peaches. A total of
160 freshly harvested peaches werefirst ranked by an expert judge into five firmby 100 untrained
nessgroups with 32 peachespergroupandthenevaluated
247
consumers by touching (no squeezing). The untrained consumers were instructed

to rank a set of five peaches randomly placed on a tray from most firm to least
firm. The ranking of the expert judges matchedwell with those of the consumers
(19).
Perhaps the human touch represents the best sensation
of firmness, being
nondestructive and utilized at the final point at which consumers may accept or
reject a product based on the combinationof appearance and force (by touching).
However, it is subjective, labor intensive, and slow.In addition, it is good mainly
for products that have thin peels and those with uniform internal structure.
2. Based onForce-Deformation
Mechanical force-deformation is the conventional means of evaluating the firmness of fruits and vegetables.The instruments measure either force or deformation
or both (dynamic force-deformation). For nondestructive firmness evaluation, the
force applied needs to be at a level that results in no damage to the product.
Mehlschau et al. (20) measured the maturity of pears nondestructively using two
19 mm diameter steel balls to apply a constant force against the opposite sides
of the fruit.
A portable device developed by Timm et al. (21) employed two parallel
surfaces to slightly compress the fruit to measure the firmness of sweet and tart
cherries, blueberries, and strawberries. The device was mounted on an aluminum
frame on which a stepping motor (incremented 0.1 mm per pulse at a constant
rate of 8.85 m/s) was used to control speed. A computer was used to control the
stepping motor and to acquire data from a load cell.
it
Mizrach et al. (22) developed a mechanical thumb device and used
to measure force-deformation of tomatoes and oranges. The method is based on
creating a slight elastic deformation of the fruit peel using a spring-loaded pin
3 mm in diameter. Two models of this device were tested, one operated
in a
go-no-go mode in which a fruit peel deformed as the pin penetrated into the
fruit and caused the movement of a pivot arm against a spring adjustedby a load
presetscrew. An adjustable small movement (e.g.. 0.2
mm) of the pivot arm
is adactivates a microswitch, thus preventing damage to the fruit. The spring
justed to a certain cut-off stiffness value that is equivalent to the stiffness of a
fruit of desiredfirmness.Fruit is consideredfirmwhentherecordedforce
is
greater than the predetermined spring load. The other mode of the device allows
continuous deformation of the peel with a pin connected to a flat spring to which
a strain gage is bonded. The deflection of the flat spring under load indicated
fruitfirmness,withfirmerfruitsshowinghigherdeflection
of thespringand
higher output voltage from the strain gage.
The two sensing devices described
above were placed on a sorting line and tested with oranges and tomato fruits
248
Hung et al.
with good results. The on-line system was implemented onan endless conveyor,
which moved the fruits towards the sensing
device attached to a pivot arm that
was free to move up or down depending on the fruit size.
Takao and Ohmori (23) implemented a similar concept an
foroff-line computer-automated system using either deformation for a preset force or maximum
on kiwifruit and melon.
force for a preset deformation. Tests were conducted
They used different compressing plunger shapes and diameters to accommodate
different fruit shapes and diameters, respectively. The plunger speed was set
at
50 mm/min and the plunger diameter ranged from 8 to 12 mm.
to nondestructively
Botta (24) used a unique force-deformation concept
measure the firmness (and resilience) of Atlantic cod fillets, which could also be
adapted for fruits and vegetables. A portable instrument was developed to apply
a force of 100 mN at 20 mm/s to just touch the surface of the product placed
on a rigid flat plane, using a 2.5 cm diameter probe.The force was then increased
Di) the prodto 4.9 N, and the probe was made to depress (deformation distance,
uct for one second. The force was then removed and the product was allowed
D,, was measured. A texture
to rebound for one second, and the rebound distance,
(firmness and resilience) index (D,/D,) was then calculated.
Davie et al. (25) used a digital micrometer for nondestructive measurement
of kiwifruit firmness. A 5 mm diameter hemisphere probe was attached near the
end of a horizontal lever such that it could depress a fruit placed under it. The
mass of the lever and its assembly on the product side are counterbalanced using
two weights (0.55 kg for coarse adjustment and 0.031 kg for
fine adjustment)
attached to metal rectangular channel made to slide along the part of the lever
on the opposite side of the cross bar. A digital micrometer was mounted between
the probe and the central pivot (micrometer spindle was 217 mm from the central
pivot and 63 mm from the probe), with its tip touching the top surface
of the
lever. The resting forceof the probe on the fruit surfacewas adjusted to a slightly
positive value using the coarse andfine weights. After zeroingof the micrometer,
a mass of 0.1 kg was applied on the probe manually and the deformation was
(SC) was calcurecorded every 5 seconds for 30 seconds. The softness coefficient
lated as the slope of deformation versus natural logarithm of time. Tests with 28
fruits were correlated with Effegi penetrometer firmness (f) using the following
equation:
where k, and k2 were coefficients.In a second prototype, they (25) used a digital
micrometer fitted with a motorized system to raise or lower the probe ( 5 mm
diameter hemisphere). After contacting the fruit surface, the micrometer was preset to zero and a mass of 0.1 kg was applied. A computer and data logger were
Methods
Firmness-Measurement
249
used to acquire a burst of three micrometer deformation readings per second for
5 seconds.
Dawson (26) described a nondestructive firmness tester for kiwifruit based
on the displacementof a small mass pushed against the fruit by
a spring. A device
for the measurement of berry firmness was reported by Armstrong et al. (27). A
230 mm diameter turntable that had 25 spherical indentations was driven
by a
stepper motor with precise 0.13 mm/step. A fruit wasplaced into each indenture
that had a small hole in the center by which a vacuum was applied to ensure the
to the stepper motor had a plate by
berry did not move. A load cell connected
which compression was applied to the fruit and the load cell signal after amplification was digitally convertedby a data acquisition system. All 25 measurements
could be completed in one minute.
All of the above-discussed methods directly measure the force and the corresponding deformation. However, Abbott and Massie (28) developed a system
to apply compressive force dynamically to apple and used the dynamic forcedeformation data in the range of 40-440 Hz to obtain a predictor of firmness.
A 1.25 cm diameter force transducer was statically loaded against the apple with
8.8 N. A two-channel dynamic signal analyzer generateda sine wave sweep(40440 Hz), which drove an electrodynamic vibrator through an amplifier. A signal
analyzerthencapturedthetimedomainsignalsfrombothforcetransducers
attached to an accelerometer, and the frequency response (ratio of output (force)
to input (accelerometer)) was used to calculate dynamic force/deformation from
each test.
The main advantage of force-deformation methods is that results give a
direct measure of product firmness. However, most direct contact methods have
limited speed and require mechanical contact with the product, thus increasing
the chances of cross-contamination of the bulk.
3. Based on Impact
Two general concepts havebeen adopted by different researchers, who have measured the firmness of fruits and vegetables
by the impact phenomenon. In the
first mode, the fruit is allowed to fall through space and impact a rigid surface,
often fitted with a force transducer. In the second mode, an impacter is made to
strike the product. Variations of these two concepts have also been explored by
also found that the impact of a fruit
some researchers. Many researchers have
on a rigid surface can be closely modeled by the impact of an elastic sphere and
that the firmness of a fruit could be obtained based on the impact force response.
Delwiche et al. (29) sorted peaches and pears into hard, firm, and soft categories by analyzing the force from the fruit impacting a plate supported by force
a rigid steel
transducers. The impact force was measured when fruit impacted
plate measuring 50 mm X 50 mm X 9 m m to which a piezoelectric force trans-
250
Hung et al.
ducer was attached. The transducer output was sampled at 1 1.62 kHz by a spectrum analyzer.
Technological advances have enabled greater ease in acquiring, analyzing,
and modeling of impact data required for improved sorting systems. Lichtensteiger et al.(30)investigatedimpactresponseforsortingtomatoesandother
to controlthe
sphericalviscoelasticobjects by incorporatingavacuumpump
release of a sample onto an aluminum plate (3.75 cm
in diameter and 0.32 cm
thick) glued to the top of the force transducer to increase the impact area and to
lower the natural frequency from 70 to 10.5 kHz.
Rigney et al. (31) developed an experimental machine to orient peaches
for on-line quality measurement as
an aid to more consistent results where a
precise presentation of the fruit is needed. Also, to overcome the variability aris(32) developed
ing from the influenceof angle and location of impact, Ozer et al.
a method based on multiple impacts. In their extensive study of nondestructive
firmness measurement of cantaloupes, Ozer et al. (32) used a drop test from a
fixed height of 2 cm to an impact surface. The impact response data of the fruit
was captured by a force transducer at a sampling rate of 10 kHz. The same fruits
were subjected to four multiple impact tests starting at the groundspot and moving
along the equator to locate two other points, and the fourth was at the blossom
end.Foreachimpact,theimpactparameter,
FT, wascalculated to represent
firmness:
where P, is the amplitude of the impact response (V), T,, is the starting time, and
T,, is the end time of the impact, s.
Theconcept of multipleimpacts on thesamefruitwasautomated
by
on afruitfirmnessgraderforkiwifruit(Softsort
SchaareandMcGlone(33)
grader) based on four impact stages on specially designed concave steps as the
into the receiving
fruit tumbled down an inclined chute. Unsorted fruit was fed
conveyor, which moved the fruits through a single lane over inspection rollers
where an operator removed fruit with external defects. The whole (unblemished)
fruit thentumbleddownaninclinedchutecomprisingfourseparateconcave
steps, each vibration-isolated (with sheets of rubber foam) from the grader frame.
Each concave step served as an impact sensor formed by folding an aluminum
sheet over a heavy steel base.A piezoelectric film sensor was glued to the underside of the aluminum step to sense the flexure of the aluminum during fruit impacts. The fruits were released at intervals to ensure that only one fruit was on
the sensor at one time. The voltage producedby the piezoelectric film as the fruit
impacts on a sensor is amplified and processed by a microcontroller attached to
each of the four sensors. The four microcontrollers transmit firmness measurements to a dedicated microcontroller, which runs the grader.
Methods
251
Bajema and Hyde (34) provided a guide for determining the suitability of
an impact-based nondestructive method for fruits and vegetables. They developed
an instrumented pendulum (3 m high) to study impact phenomena. The device
was equipped to record force and contact area profiles during impact at the rate
of I O kHz onto a flat rigid anvil. The anvil was fitted with a force transducer
and with a contact areasensor. Two precisely spaced infrared beams (beam interrupted by the sample) allowed approach and rebound velocities to be measured.
An accelerometer attached to the trailing end of the product measured deceleration during impact and free vibrations during rebound. The impact profile data
were sampled at I O kHz per channel. Using the setup, they developed schemes
for impact sensitivity for fruits and vegetables. Operating procedures for single
impacts, constant height multiple impacts (for determining bruise resistance), and
(to determine bruise threshold height) were
increasing height multiple impact
discussed.
Chen et al. (35) used an impact-testing system fitted with an accelerometer
and electromagnet and a computer-based data acquisition system to measure the
firmness of pears. Based on the theoryof elasticity, they determined that the time
(t) to reach peak force was given by
DIV
t =
1.47
where D is maximum deformation and V is the relative velocity

assuming negligible gravitational effect during impact.
The peak acceleration (A) of the impactor was expressed as
(3)
of approach
A = Flm,
where F is the peak impact force and m, is the mass of the impactor. The impact
force F depends on impact velocity V, mass of the impactor, mass of the fruit,
fruit modulus of elasticity, Poisons ratio of fruit (0.49), radii of curvature of
fruit (35 mm), and impactor (9.5 mm). The firmness of the tested product was
defined as the ratio A h .
The impact method can be high speed. and some studies have demonstrated
good firmness sorting results; e.g., Chen et al. (35) achieved good sorting ability
on a prototype packing line
by measuring the deceleration of a low mass impactor.
However, a major problem associated with the measurement of firmness by the
impact principle is the uncertainty of the angle and location of the impact on the
fruit surface, especially when such surfaces are irregular.
4. Based on Impact-Rebound
In the preceding section, it was shown how some researchers have used the
reit, to directly measure firmness
sponse of a product to an impact either on, or from
of the product. A unique variantof this principle utilizes the characteristic trajec-
252
Hung et al.
tory information generated when a product is allowed

to rebound following an
impact. This principle of impact-rebound has shown some promise for firmness
evaluation.
A rotating steel drum method based on impact-rebound has been used for
(36). Inthesestudies,oranges
separatingsoftorangesfromundamagedones
moved in lanes created by 0.3 cm thick, 5 cm high dividers on a belt conveyor
inclined at 15" to the vertical. At the top of the conveyor, the fruits were allowed
to descend in a freefall on to a rotating (30-130 rpm) steel drum and bounce
of different firmness
into a horizontal conveyor padded with soft rubber. Oranges
grades assumed different trajectories, and the horizontal bounce position on the
conveyor provided a means of separating unwholesome or cull fruit from the
(37)
bulk. A patented design of this concept for separating clods from onions
using either forward- or backward-bouncing modes was developed. The separation efficiency (E) was determined as the product of recovery and rejection as
given below:
where 0, is the number of onions in product exit, 0, the number of onions rejected, C, the number of clods in reject exit, andC , the number of clods remaining
with product. They also used a similar system to separate clods from potatoes
(38).
Although the impact-rebound method has the advantage of high capacity,
it is unsuitable for products having a wide variation in shape. In addition, it is
unsuitable for delicate products.
5. BasedonVibrationalCharacteristics
The vibrational characteristics of fruits and vegetables depend on their elastic
modulus (firmness), mass, and geometry.A form of impact-rebound is generated
when fruits are placed on a vibrating surface, and this principle has been studied
as a means of sensing firmness. In a review of nondestructive firmness measurement, Chen and Sun (12) found that prototype equipment had been built by Bower
and Rohrbach (39) for sorting blueberries and grapes by low-frequency vibration
(200 Hz). The fruit was placed in an L-shaped trough, one side of which was
vibrated at a fixed frequency and amplitude of 200 Hz and 0.3 mm, respectively.
Firm fruit bounced outof the vibrating trough and intoa collecting trough, while
softer fruit was conveyed along the trough by the vibration.
The problem of mechanical contact between the fruit and sensor appears
to have beenaddressed in a recentreport (40) on a uniquecommercialfruit
firmness sorter for apples, nectarines, and kiwifruit. The machine is comprised
of the fruit
of a conveying system, which allows nondestructive physical contact
ethods
253
and a sensor finger while the fruit was on-the-go at speeds of 2.5-7.5 fruits/
s per lane. An electrodynamic shaker wasused to vibrationally excite the bottom
part of the fruit. The root mean square (RMS) of the input signal was measured
in the shaker head while the RMS
of the output signal (fruit response) was picked
up with a miniature accelerometer attached to the sensor finger contacting the
top of the fruit. A firmness index (PFT) was defined as
PFT = X,/(X,,
X,)
where X,, and X, were the input and output RMS signals. Larger PFT indicated
firmer fruits, because softer fruits tended to attenuate the input vibrational energy,
while firm fruits pass a larger portion of the input vibrational energy.
Vibrational methods have the advantage of firmness evaluations based on
a response from the whole item. However, they can be limited by the influence
of geometrical differences, such as size. There is also therisk of stimulating
whole-fruit softening damage, depending on the regime of vibrational frequency
and amplitude employed. Also, a technical problem that needs to be solved is
theabilitytomaintaingoodmechanicalcontactbetweentheproductandthe
sensor, since this is necessary to faithfully transfer the product response signal
to the sensor.
6. Based on Soft Touch

Difficulties and the destructive nature
of the MT method of firmness measurement
have led researchers to develop nondestructive methods for replacing the technique. Perry (41) designed and tested a nondestructive firmness laboratory device
to releasecompressedair(62mPa)
forpeaches. A solenoidvalvewasused
the fruit being
through an air hose into two pressure cells on opposite sides of
tested. The pressure cell was made with a transparent Plexiglas cylinder (6.4 cm
a silicone rubber diaphragm at
internal diameter, 0.32 cm wall thickness) with
the opposite end of the cylinder for holding the fruit and preventing air leakage.
Adialindicatorwaspositionedinsidethecylinderwithitsstemprotruding
through the diaphragm and resting lightly on the fruit surface. Test data were
acquired by releasing compressed air into the cells for a specified time interval
and recording the deformation (indicator of firmness) of the fruit as indicated by
the dial gages.
Also, Thai (42) developed a novel nondestructive firmness detector, which
used a soft foam transducer/tactile sensor (STTS) to evaluate apples, peaches,
and tomatoes. The fruit was dropped onto theSTTS sensor (thickness 1.27-6.35
mm) from a height of 5 cm. The change in electrical resistance at contact was
measured by putting a known and fixed resistorin series with it. Signal collection
(acquisition rate of 8 kHz) was triggered by an near-infrared (NIR) beam gener-
254
Hung et ai.
ated by light-emitting diodes placed about 6.35 mm above the sensor. Firmness
was determined by the change in electrical resistance.
The main advantage of the soft touch methodis that it is intrinsically nondestructive. However, it is slow and difficult to integrate on-line, and in addition,
different product size may affect accuracy.
C. NondestructiveMethodsBasedonSecondary
Properties
Themeasurement of secondaryproperties (not force-deformation)greatlyincreases the number of possible approaches. Most of the research has been on
resonant frequencies, acoustic response, and ultrasonic methods. A major advantage of firmness methods based on secondary properties is that they evaluate the
whole product. Some applications can also utilize remote sensing with no physical contact with the product. However, general limitations include the fact that
some of the properties may not phenomenally relate to firmness, and i n some
methods product size may affect the readings. Also, in some applications the cost
of equipment may be high.
The challenge with the firmness-sensing methods using secondary properties is either to develop devices that are independently accurate or to develop
methods having very high correlations with the results obtained with force-deformation relationships of the product that are important to buyers and end users.
Also, the speed of methods that require mechanical contact or signal processing
A technical complication
should be addressed for on-line sorting applications.
of NMR is the need to avoid metallic objects near the vicinity of the sensor.
1. Based on Acoustic Response

Muramatsu et al. (43) used acoustic vibration t o measure kiwifruit firmness using
a pulsed sound, one-half wavelength at1 KHz. The sound generatedby a function
synthesizer was transmitted to the lower surface of the fruit by a 3 cm diameter
speaker. A small contact microphone was attached to the opposite surface of the
fruit. To facilitate sound transfer, a small amount
of clay (or athin sheet of rubber,
I mm thick) was placed directly between the fruit and speaker. The transmitted
wave was monitored with a storage oscilloscope and displayed together with the
imposed wave. The extent of wave displacement was usedto estimate the elapsed
time for sound transmission through the fruit, which was proportional
to firmness.
The elapsed time for sound transmission through the fruit increased as fruit became softer.
Sugiyalna et al. (44) developed a portable firmness tester for melons based
on the acoustic response of the fruit to an impulse. They used a mechanical impulse generator-a computer equipped with an analog-to-digital(A/D) converter
Methods
255
to measure sound transmission velocity using two microphones placed

on the
melon surface 16 mm apart. The sound signals from the two microphones were
amplified (gain of 50) and transmittedto the A/D converterat sampling frequency
of 70 kHz per channel. To ensure high values of signal-to-noise ratio, a long
impact rod isolated noise due to friction between the spring and the barrel of the
impacter was used. The rod and barrel were made with resin (polyacetal), which
reduced the amount of noise generated when the microphones were placed next
to the impact point. The average time difference between the two sound signals
(t,l) from the microphones was determined;
where id is the number of data points, t, ( = 14.2 ps) is the sampling period (70
kHz sampling/channel), and 0.5 was added t o account for the difference in sampling time between the two channels caused by the multiplexer of the A/D converter. Sound transmission velocity (V,l) was expressed as;
Vd = d/t,
where d is the distance between the microphones. The average of five measurements of Vd from different locations on each fruit was determined and used as
the nondestructive firmness.
2. Based on Sonic and Ultrasonic Technique

Finney (45-47) demonstrated that subjecting peaches to a random-noise generator or a beat-frequency sine-wave generator and that measuring the amplitude of
vibration of the fruit can provide data for sensing the fruit firmness. Amplitude
measurements on the blush and nonblush cheeks of each fruit were made over
a frequency range of 250 to 10 kHz. The second and third natural resonances
(fn=2,1r3)
correlated well with fruit MT firmness penetrometer tests.
Abbott and Liljedahl (48) conducted sonic measurements on apple fruits
to determine their firmness. Intact apples were mounted horizontallyusing floral
to an electromagnetic vibrator. An
clay on a small aluminum pedestal attached
electronic signal equivalentto a 5 to 2 kHz sinusoidal scan from a signal generator
was fed to a power amplifier to drive the vibrators electromagnet. The resulting
pulse caused input vibrations of constant acceleration within the imposed frequency domain, which was measured with an accelerometer weighing 0.5 g held
against the opposite sideof the fruit with an open cell polyurethane sponge under
an arm. The amplitude and frequency spectrum contentof the accelerometer signal was extracted by FFT with a resolution of 5 Hz. Apple stiffness coefficients
defined as fm and fm3 (m = mass and f = frequency, calculated from either
the second or third resonances f2 and fJ were used to indicate firmness. They
Hung et al.
256
found that frequencies of resonant modes 2 and

3 are closely correlated (r =
0.95) and seldom varied by more than 3%.
Firmness was also evaluated by the ultrasonic technique at an appropriate
frequency range and power level (49) to evaluate different fruit-tissue specimens
(50,51) and whole melon and avocado (49,52). Numerous researchers investigating fruit firmness by frequency analysis have used the stiffness coefficient (fm
or f m23)to represent the firmness. However, Galili et al. (52) developed a new
frequency-related firmness parameter, F,, which is the centroid of the frequency
response. Mathematically, F, was expressed as
where f, and H, are the local values of frequency and amplitude of the signal,
respectively, and n is the total number of data points. The meritof this parameter
is that it does not depend on the specific resonant frequency, but uses the entire
spectrum of frequency response.
Mizrach et al. (50) measured nondestructive firmness of mango fruit by
A high-power, lowmeans of ultrasonic probes in contact with the fruit peel.
frequency ultrasonic pulser-receiver, a pairof 50 kHz ultrasonic transducers, and
a microcomputer and data acquisition system were used to generate the signal and
to collect data. The ultrasonic head with a transmitter-receiver system allowed
transmission and reception of ultrasonic signals, which passed through the peel
and the fruit tissue next to the peel. Exponential-type plexiglass beam-focusing
elements were used to reduce the beam diameter of each transducer on the fruit.
(A, dB/
They found that firmness (FR, N) can be calculated from attenuation
mm) as
FR = 48.21A2 - 408.71A
+ 878.16
(10)
3. Other Methods
Recently, nondestructive texture measurement of kiwifruit, pear, and peaches by
(LDV) hasbeenreported
(53). Individualfruitwas
laserdopplervibrometer
placed on a vibration generator, and a small amount of clay was applied directly
between the fruit and vibrator stageto ensure integrity of vibration transmission.
The vibrator generated inputs to the fruit in the frequency range of 5-2000 Hz
at 0.5 Hz increments. A laser beam reflected from the fruit surface to the sensor
head was used to measure the vibration at the upper surface of the fruit. A fast
to identify the phase shift between
Fourier transform (FFT) analyzer was used
ethods
257
the input and output of the vibrational signal. A phase shift at either 1200
or
1600 Hz was used to indicate firmness. Also, Muramatsu et al.
(54)used LDV
to measure the firmness
of kiwifruit and apples. They found that using the accelerFFT
ation output as reference signal gave more accurate results than using a
output signal as in their earlier study.
Some research has been done relating firmness to NMR data
of fruit. NMR
methods are based on a processof magnetic relaxation following external excitation of a fruit. The signal induced in the magnetic coil decays exponentially with
time. The relaxation data could be expressed in two forms, butmore
the important
mechanism is known as transverse relaxation (55).
In a recent review article, Clark et al. (56) gave an overview of magnetic
resonance imaging (MRI), including NMR applications in postharvest handling
and quality evaluation of fruits and vegetables. Kim et al. (55) investigated the
feasibility of using magnetic resonance imaging
to determine the firmness of
tomato. The principle of these techniques is based on many proton-containing
species, in particular hydrogen atoms in fruits and vegetables, which are capable
of absorbing radio frequency energy.In general, fruits were subjectedto magnetic
to decay exporesonance, and the signal induced in the receiver coil was allowed
nentially with respect to time. The relaxation allowed the spins to return to their
equilibrium levels, releasing their surplus energy into the surroundings.
The spectral data can be analyzed by two relaxation mechanisms. Longitudinal relaxation
(spin lattice relaxation) was expressed by (55) as
where, MY is the initial amount of longitudinal magnetization, M(t), is the magnetic moment at a certain time t, and T , is the spin-lattice time constant. Transverse relaxation (spin-spin relaxation) is given by
where Mf is the initial amount of transverse magnetization, M(t), is magnetic

(55)
moment at a certain time t, and T2 is spin-spin time constant. Kim et al.
measured the Ti and T2time constants usinga 0.6 T machine and bird cage coil.
The terms l/Tz and In (Tz) were used to represent the fruit firmness.
Another unique method for fruit firmness measurement involves the useof
delayed light emission (DLE). In a review by Gunasekaran (57) it was reported
of apricots and
that DLE has been used successfully to measure the firmness
of the method include
papayas (58). However, many factors influencing accuracy
wavelength of excitation, excitation intensity, time during and after excitation,
length of dark period during which DLE decays, sample size and surface characteristics, temperature, and chlorophyll content. Also, NIR (near-infrared) spec-
Hung et al.
258
troscopy in the range of 380-1650 nm has been used to nondestructively estimate

the firmness of apples (59).
D. FactorsAffectingFirmnessMeasurement
1.
Temperature
Often researchers simply conduct firmness tests at prevailing room temperature

conditions or, at best, stipulate the temperature at which the measurements should
be made. However, published data show that the temperature of fruit has a significant effect on the measured firmness. Bourne (10) proposed that temperature
effect can be characterized by a firmness-temperature coefficient (KIT, %/C):
where, f , and f2 are firmness at T, and T2, the lowest and highest temperatures,
respectively. Jeffery and Banks (60) adopted this concept and obtained equations
to evaluate the effect of temperature on the firmness of kiwifruit measured destructively at 0 and 20C. They also found that fruit origins had no effect on the
relations between KITand f , or f2, suggesting
that within thescope of experimental
to be used for fruits from a variety of
error, the models may be robust enough
sources. In addition,temperaturesbelowzeromayyieldfirmnessvalues
that
differ significantly from those measured
at higher temperatures. Jackman and
rigidity of refrigerated
Stanley (61) found that the increased shear strength and
tomatoes may be dueto a chilling injury mechanism in which the cell wall structure binds together firmly.
2. Composition and Structure

Variation of mechanical properties within a fruit, generally determined with tissue samples, may also be observed due to internal variation of the fruit tissue.
to measure
The importance of thevariationmaydependonthemethodused
et al. (30) found that shelleffect may have
firmness.sinceLichtensteiger
relative significance on a measurement. In the case where the core/inner layers
are softer than the relatively thin external shell, a sudden change in the slope of
the force-time curve occurred shortly after contact with the object. Lichtensteiger
et al. (30), in their impact force analysis study on tomatoes and blueberries from
drop height of 100 mm, found that the skin layer of the fruit may create what is
generally known as a shell effect. They confirmed that this effect
is more
prominent when the internal structure is softer than a relatively thin external shell
and less prominent when the internal elements are stiffer than the shell.
259
The shell effect was also found by Ezeike and Otten (62) in cotyledonous
seeds where an air layer separates the outer cotyledon from the seed. Abbott and
Lu (63) found that the stress, strain, energy at failure, and Youngs modulus of
apples varied with orientation at point
of measurement and location within the
fruit. Youngs modulus was higher in radial samples than in tangential and vertical samples. They also found that fruit variety was a contributing factor. For
example,DeliciousapplesshowedgreatervariabilitythanGoldenandRome
apples. Similar findings were reported by Maness and Brusewitz (64), who mea8
sured peach firmness using an Effegi penetrometer equipped with a standard
mm diameter probe, on samples taken from inner, middle, and outer regions of
the mesocarp at four angular positions around each peach half. Inner mesocarp
was firmer than outer mesocarp. Variations were also found between middle and
outer regions of the mesocarp and were cultivar dependent. Firmness was found
to decrease longitudinally from the stem endto the blossom end and latitudinally
from the suture to the cheeks.
Weexpect that firmnessmethodsbased on sensingtheresponse of the
outermost layers of fruit may be subject to the shell effect if the fruit has a structure with distinct outer and inner tissues. On the other hand, firmness methods
that excite the whole fruit (e.g., acoustic methods) are less subject
to the shell
effect.
3. OtherFactors
In their study of effects of mass and drop heighton kiwifruit firmness by impact,
(a measure of impact
McGlone et al. (65) developed equations for dwell time
duration) and peak force and found that the firmness parameter Cz (= f&, f,, is
the peak force and t, the dwell time) was independent of mass of the fruit. They
also found that fruit was not damaged even with repeated impacts
at 50 mm. Chen
and De Baerdemaeker (66) conducted studies with apples and Litchtensteiger et
al. (30) with spherical viscoelastic objects and tomatoes to optimize the parameters involved in impact firmness testing. These parameters also affect the performance of different firmness methods, which include impactor mass, fruit mass,
internaUexternal properties, maturity/damping, calculation method,
initial contact velocity, flat contact surface, elasticity of material, and stiffness factor.
111.
NO MECHANICALCONTACTWITHPRODUCT
A.
Concept
A nondestructive, noncontact method for measuring the firmness

of food and
other products was developed at the Universityof Georgia by Prussia et al. (67).
The laser air-puff food firmness detector uses a brief puff of compressed air to
260
Hung et al.
deform the product surface about one millimeter.

A laser displacement sensor
provides a rapid and accurate measurement of the deformation. At a fixed air
of a firm product is less than for a soft
pressure, the maximum deformation
product. When changing to a different type of product with a different firmness
range, the pressure is easily adjusted to the level needed to causea deformation
slightly less than necessary to cause damage. For example, a pressure setting of
70 kPa is common for soft peaches, while 400 kPa is needed for evaluating of
apples.
B. Original Design
Extensive tests with the first prototype verified the validity of the laser air-puff
in its accuracy and repeatability for
a wide
concept and provided confidence
range of fruits, vegetables,and other products (68). Compressed air was supplied
through a pressure regulator to a tank located directly over pilot
a
actuated valve
with an attached 9.5 mm diameter nozzle (Fig. 2). A laser displacement sensor
was mounted at a 30" angle to the nozzle and aimed at the top of the product
that was supported under the nozzle in a specially designed holder with vertical
adjustment. A timer circuit controlled the duration of the puff (valve opening)
after a switch activated the circuit. The amount
of deformation caused by the
compressed air was measured using a laser displacement sensor and recorded
with a digital oscilloscope.
A second pressure regulator on the main air supply assured that consistent
pressure was delivered to the pilot side of the pilot actuated valve (ensuring repeatable speed of opening). A large tank (8 L) was used to minimize the drop
in pressure while the valve was open for 200 ms to produce the puff of air. The
laser displacement sensor (Keyence model LB 041) was mounted 40 mm from
the product surface (stand off distance). The laser beam was aimed at the center
of the deformation when the product was 2 cm from the nozzle. Vertical adjustment was provided for the product holder to assure that the top surface
was at
the set point regardless of variations in product size. Spherical objects were supported by three cups about1 cm in diameter and with a radiusof curvature slightly
larger than the object. The cups were supported on a ball-and-socket joint that
allows movement of the cups for accommodating products of irregular shape.
Several stepswere required to obtaina firmness measurement. Product was
placed in the holder, positioned under the nozzle, and the height was adjusted
until a voltmeter connected to the output of the laser displacement sensor was
zero (indicating the nozzle-product distance was2 cm). The pressure in the tank
was then set by adjusting the pressure regulator until the desired pressure was
reached as indicated by a precise digital pressure sensor (1.7 kPa resolution). The
switch for the solenoid valve was then pressed when the digital pressure sensor
indicated the desired value.
261
Fig. 2 The original laser air-puff food firmness detector.
Output voltage from the laser displacement sensor was

5 5 V, corresponding to the displacement 2of5 mm. The typical voltage trace on a storage oscilloscope rapidly increased as the
puff of air caused the product to deform, decreased
slightly after reaching a maximum, and then rapidly decreased when the
puff
ended. Peak deformation was found by subtracting the final displacement from
the maximum (the final displacement accounted for any movement in the holder
that occurred after the initial value). Mean deformation values had to be estimated
4 mV noise representing
from the oscilloscope trace, which typically had about
a displacement of 4 pm and occasional spikes.
Figure 3 shows that plots of maximum displacement versus tank pressure
for measurements made on a rubber ball were extremely linear (R2= 0.997).
Hung et al.
262
3000
A
E
3 2500
*
C
y = 6.51ffiX - 295.78
2000
0
5m 1500
6
5 1000
-E
i!
500
o f
0
100
200
300
400
500
Tank Pressure (kPa)

Fig. 3 Effect of puffof air pressure on maximum displacement of a rubber ball.
Plots obtained for fruits and vegetables were also linear. Repeatable results over
a wide range of products and conditions indicated that the concept and design
was successful.
to holdthemainstructure,
A 1 m X 2 mtableorbenchwasrequired
oscilloscope, laser control box, electronics box, and volt meter. Although it was
it was somewhat difficult at 90 kg and with
possible to move the equipment,
several electrical connections. Due to the large size and complicated operating
procedures of thefirst prototype, theunit was redesigned to facilitate easier operation.
C. Improved Design
Figure 4 shows the improved prototype design of the laser air-puff unit, which
is about the size of a tower computer. The componentsof the system are similar
to the original prototype except that they are smaller in size and with electronic
interfaces. A notebook computer is used for controlling a pressure regulator and
a solenoid valve, for acquiring data from the laser displacement sensor, for anain a spreadsheet format. The software is
lyzing the data, and for storing data
a visual programming package called LabView from National Instruments. A
PCMCIA card interfaces the computer with the hardware.
The fruit sample holder shown in Fig. 4 is a new design that uses three
balls to support the product rather than three cups as in the first prototype. The
Methods
Firrnness-Measurement
263
Fig. 4 The improved prototypeof the laser air-puffuses anotebook computer to control
the measurements and to analyze and store the data.
cups had allowed some movement, which was only partially compensated by
for
calculating maximum displacement from the final rather than the initial displaceof movements. The purpose of using the balls is to allow a repeatable amount
ment rather than trying to eliminate all movement of the product.
The advantages of the laser air-puff approach include no contactwith the
product, high speed, and good accuracy. However, it is limited
by its cost (about
$lO,OOO) and the need for an improved sample holder.
IV. TOOLSFOREVALUATINGFIRMNESSMEASUREMENT
Nondestructive methods for detecting firmness are typically evaluated
by comparing results with the destructive MT test. For over 75 years the MT test has been
the standard for measuring firmness. One of two plungers (7.9 or 1 1.1 mm in
diameter) with a specific convex tip is inserted into the product to a depth 7.9
of
mm. The maximum penetration force is the value that represents the firmness.
Results should specify the plunger to be used and the maximum force (not pressure).
Hung et al.
264
Comparing a nondestructive measurement of firmness with the results

of
the destructiveMT test has obvious difficulties. First, different physical phenomena occur during a nondestructive, viscoelastic deformation comparedto the destructive rupturing of tissues and cells for the MT penetrometer test. Comparing
results when different physical properties are evaluated
is somewhat like comparing apples with oranges. For example, McGlone et (69)
al. observed some anomalies in firmness of kiwifruit measured nondestructively by impact force analysis
of I O mm,
during medium to long periods of cool storage from a drop height
compared to firmness determined by penetrometer method. They found that the
softening curve of kiwifruit stored at refrigerated temperatures showed an unexpected plateau and/or rise during medium to long periods of cool storage, but
not during room temperature softening (between 15 and 25C). The anomalous
behavior was attributed to the fact that impact force response was closely correlated with whole fruit stiffness, a fruit property that
is different from the flesh
rupture force measured by the penetrometer. Whole fruit stiffness
is usually measured by parallel plate compression and can remain constant for long periods, or
even increase slightly, while the flesh rupture force measured by
MT continues
to decrease.
A second difficulty is that the MT test is a measurement method that has
about the same correlation with established engineering properties as other instruments. Thus, comparison of the instrument with the MT results in even lower
correlations.
A.
Modulus of Elasticity
To avoid the problems mentioned above, one solution

is to use an engineering
property that can be calculated from different firmness-sensing methods. Fortunately, modulus of elasticity (E) can be used as such a standard. The American
Society of Agricultural Engineers (70) gave the following expression for compression test of food materials of convex shape:
where
EASAE
= modulus of elasticity, MPa
R , = maximum radius of the tested spot, m
R; = minimum radius of the tested spot, m
d
F
D
p
diameter of the indenter, m

force applied, N
deformation caused by the force
= Poisson's ratio
=
=
F, m
265
The value for E found withthe ASAE procedure can be checked with other
methods such as the removal of a specimen of known shape, which allows direct
calculation of E from cross-sectional area, force, and deformation data.
A similar equation for E was developed by Fan (68) for the laser air-puff
method as
where
ELASER= modulus of elasticity, MPa
p = Poissons ratio
P = pressure in air tank, MPa
V = maximum voltage reading from laser sensor,
mV
Extremely high correlations (R2 = 0.92) were obtained between values of

EAsAE and ELarrrfor peaches as shown in Fig. 5. The linear regression of the data
also had a slope closeto 1.0 (1.08), and the intercept was reasonable(0.241 MPa).
2.5
1.5
y 1 . 0 8 ~+ 0.241
R z = 0.Q2lS
0.5
1.5
2.5
EASAE (MPa)
Fig. 5 Correlation of modulus of elasticity calculated from the laser air-puff firmnesssensing data (ELASER) versus modulus of elasticity calculated from the deformation data
using a spherical indenter (EASAE).
(unpublished data)
266
Hung et ai.
A similar study with apples by Fan (68) also had a high correlation between the
two methods for calculating E (R = 0.91).
New Zealand researchers (65) also had extremely high correlations (R? =
0.94) for kiwifruit between E values calculated from parallel plate compression
and values for E calculated from an equation using drop height and dwell time
when the fruit impacted their SoftSense instrument. The New Zealand researchers
(65) expressed the whole fruit stiffness (E) based on the parallel plate method
(ASAE Standard S368.1) with the Instron Universal Testing machine as
where F is the compression force, r is the radius of curvature of fruit, and D is

the sample deformation corresponding to the force.
Also, their equation for Softsense (EJ stiffness calculated from Herz contact theory was
where mis the fruit mass, is

r the radius of curvature of the impacted fruit surface,
td is impact dwell time, h
is the drop height, and p is the Poisson ratio (0.49
being considered appropriate for a hard fruit).
6. RegressionAnalysis
Resultsfrom a nondestructive test areusuallycomparedwiththe
MT test or
MT or other tests on
another method by regression analysis. Results from the
products with a range of firmness values are typically plottedon the independent
axis and values from the new method on the dependent axis.A correlation coefficient is used to judge the performanceof the new method. For biological products
such as fruit, a correlation coefficient above 0.70 is considered good.
However, correlation coefficients are affected
by the number of data points,
the range of the data, and the distribution of data over its range. A computer
spread sheet program was developed for simulating correlations between the outa hypothetical new
put of a hypothetical standard instrument and the output of
instrument (71). The initial simulation had 100 data points uniformly distributed
with a range of 12 N (4-16 N) on both the x and y axes. Each data point was
given a random variation on each axis with a normal distribution having a mean
atthepointandastandarddeviation
of 1 N. Linear regression analyses gave
differentresultsforslope,intercept,andcorrelationcoefficienteachtimethe
simulation was run (much like comparisons of real instruments).
267
Typical values for R' increased from 0.9 to 0.95 when the range of the
to 16 N (4-20
uniformly distributed points was increased from the original 12
N). The typical values for R? decreased from 0.9 to about 0.8 when the ranges
in the data points were determined by using 100 normally distributed points with
a certain mean ( I O N) and standard deviation (23 N) (Fig. 6). There was very
little change in R' values (typically 0.8) when the number of normally distributed
data points was increased from 100 to 200. These results demonstratethat values
to very firm fruit in a test. The
for R' can be increased by including very soft
range used for evaluating an instrument should be the same as the range that is
important to the user.
Regression analyses do provide a useful tool for making general comparisons between two firmness-measurement methods. However, rather than summary statistics, commercial operations such as sorting fruit at packing houses
20 2 18 4 16
614
012
10
Output of Standard Instrument (N)

Fig. 6 Simulatedcomparisonofoutput of a newinstrumentwiththatofastandard
instrumentusing 100 normallydistributedpoints(mean
= 10 N; standarddeviation =
"3 N). The precision of both the standard and the new instrument was also simulated by
adjusting the position of each point in
x- and y-directions according to normal distribution
(mean = original value; standard deviation = 2 1 N). The boxes show calculations for
hit, miss, false alarm, and correct acceptance rates for an arbitrary cut-off value of 6 N.
268
Hung et al.
depend on accurate differentiation among individual items with similar firmness

values.
C.DiscriminateAnalysis
The following description of discriminate analysis shows how individual items
are evaluated for a quality characteristic. A firmness-sorting operationat a packinghouse has the purposeof removing items that are softer or firmerthan a specified level. There are only four possible outcomes for such an operation with either
accept or reject decisions (72):
Hit
Miss
False alarm
Correct acceptance
Removing a defective item (too soft

Allowing a defective item to pass
Removing of a good item
Allowing a good item to pass
or too firm)
Rates are calculated by dividing the four outcomes by the maximum number of
hit rate is the
itemsavailable in thecategorybeingsorted.Forexample,the
number of defective items correctly removed dividedby the total number of defective items in the lot. The miss rate can be obtained as the complement of the
hit rate or calculated by dividing the number
of misses by the total number of
defective items. Similarly, the false alarm rate is calculated
by dividing the number of good items discarded by the total number of good items in the lot. Again,
correct acceptance rate is the complement of the false alarm rate or the number
of good items allowed to pass divided by the total number of good items.
For a sorting operation the miss and false alarm rates are more important
than the hit and correct acceptance rates. The miss rate indicates the percentage
firm), while the false alarm
of items in the final pack that are too soft (or too
rate represents the percentage of good product that could have been sold at full
price but was discarded.
The calculations shownin Fig. 6 are examples of the simulated comparison
of the output of a new instrument with thatof an existing (standard) method with
an arbitrary firmness cut-off value of 6 N set for soft fruit. The comparison was
made by plotting 100 normally distributed points (mean = 10 N; standard deviation = kc3 N). The precision of both the standard instrument and the new instrument was also simulated by adjusting the position
of each point in both the x
and y directions according to a normal distribution (mean at original point with
a standard deviation of -C I N). The I O points to the left of the vertical line at 6
N represent soft fruit used for calculating the
hit and miss rates. The 6 points
below the horizontal line at 6 N and to the left of the vertical line represent hits.
The hit rate of 60% was calculated by dividing the number of hits, 6, by the total
number of soft fruit, IO. The miss rate of 40% is calculated by dividing the 4
269
misses by the total number possible, 10. The number of false alarms was determined by counting the points to the right of the vertical line and below the horizontal line. The false alarm rate of 3.3% was calculated by dividing 3 by the 90
goodfruit (100 lessthe 10 softones).Similarly,thecorrectacceptancerate
(96.7%) was calculated by dividing the number of pointsto the right of the vertical line and above the horizontal line, 87, by the total number of points to the
right of the vertical line than, 90.
The R value of 0.81 1 shown in Fig. 6 is reasonably high compared to
or other biological products. Similar values
many correlations involving food
were obtained each time the simulation was run. However, the miss rates ranged
2 to
from 10 to 30% for the same runs, while the false alarm rates ranged from
8%, indicating that high R2 values do not assure low miss and false alarm rates.
The high miss rates indicate that poor sorting results can occur even when
an
instrument has high values of R compared with a standard, which also has vari2 1.0 N; Fig. 6).
ability (such as MT-simulated with a standard deviation of
Thus, it is understandable why a packinghouse manager is more interested
in the number of soft fruit in a shipment than the R2 values from a comparison
of two instruments. Discriminate analysis provides practical results for evaluating
the performanceof a firmness-measurement method. However,it is still necessary
to have a standard of comparison to determine the miss and false alarm rates.
V.
PERFORMANCEEVALUATIONOFDIFFERENT
METHODS
Consumers often base their choice of product on their perception of the food
quality assessed by their senseof smell, touch, and sight. Therefore, the relationship of objective (nondestructive) methods of quality evaluation need to also
of this
closely mirror subjective evaluations made by the human. In the remainder
section. different methods usedby researchers recently to measure the firmness
of
fruits and vegetables are summarized. The information is presented in Tables I 3 showing the principle involved for each
method reported and the typeof sensor
system used to record the response. In addition, the products evaluated, measurement conditions, and performance factorsof each method are presented as well as
the stageof development of the technique, along with the sourceof the reference.
A.
DestructiveFirmnessTestingMethods
Results of firmness measurements with the MT and Effe-Gi tests were shown by
Abbott et al. (73) not to be entirely interchangeable, even though the probes and
indicated force ranges were essentially the same. Different sizes and shapes
of
Table 1 Firmness Measurement by Force-Deformation Methods (with and without mechanical contact)
Principle
Force deformation
Measurement
Deformation
Plunger
Product
Tomato
Tomato
Measurement condition
Test 4 - 5 T
92-99% RH
Test ambient
Oranges
Deformation
Parallel plate
Moment
Deformation
Puncture force
Crushing
strength
Kiwifruit
Test 0C
Kiwifruit
Store room temp. and/or low

pressure
Melon
Blueberry
Cherries
Blueberry
Strawberry
Store 2 1 "C. 57% RH Test room
Apple
temp.
Cut specimen
Pear
Apple
Kiwifruit
Laser air-puff Displacement

and pressure
Store 2C
Test 2C
Store 12C Test room temp.
Store 0C. 95% RH. Test 24'C
Peach
Apple
Store 1C
Test room temp.
Test room temp.
Test room temp.
h)
Performance of device
R' = 0.903 (device vs. penetrometer) cv. 0.17-0.28
Accuracy 100% (separation red from
green tomatoes)
Status
Ref.
Research
83
Research, on-line
prototype ( I 0
fruitds: 3.5
fruitds)
22
Accuracy 86% (device vs manual):

cv. = 0.016-0.043
R' = 0.99 (device vs. Effegi penetrometer)
R2 = 0.93 (device vs. fruit pressure
tester value)
Research
25
Research and
commercial
23
Commercial
27
Total misclassification rate of 30%
Research
21
cv = 0.03 to 0.133 (parallel plate)
Research
74
Research
17
Commercial prototype
Patent research
prototype
7
68
0.98 (devise vs. berry condition)
-l
0
cv. = 0.70 to 1.88 (puncture)

R' = 0.98 (crushing strength vs. soluble solids)
R2 = 0.86 (crushing strength vs. Effegi penetrometer)
R' = 0.15
R 2 = 0.91 (modulus from laser v s .
modulus from spherical indenter)
RH = relative humidity (5%); Store = stored at indicated temperatudcondition: Test = temperature during test: temp. = temperature.
e
a
Methods
271
the two instruments as well as spring characteristics were thoughtto be responsible for the differences observed. In addition, operators reacted differently to the
machines in anticipation of sudden failure. Lack of sphericity or roundness of
in alignment of the fruit, the tester, and
most fruits also presented difficulties
compression surface. Oftenthe variability in measured results occurred by removing the skinto unrepeatable depths during sample preparation, since subcutaneous
cells are smaller than those nearer the fruit center. Penetrometers in general are
unable to assess flesh properties as a function of depth, and this becomes a problem for fruitsin which internal changes can occur, for example, columella softening in kiwifruit or central softening in mangoes.
B. NondestructiveForceMethods
Methods for measuring the relationships between the deformation resulting from
an applied force give a direct measure of firmness because the slope of a forcedeformation curve can be used to calculate the modulus of elasticity or stiffness
of the product (Table 1). Mehlschau et al. (20) found a curvilinear relationship
between firmness values andMT firmness. Typical characteristic curves for apple
and pear were analyzedby nonlinear rheological models by Wan et al. (74) using
apple and pear. Although the sorting rate was high (four fruitds), more accurate
results were obtained for softer fruits than firmer ones. This
is considered a major
limitation of this system according to the study of fruit classification with nondestructive firmness index based on either weighted grade purity or contamination
by Peleg (40).
Ozer et ai.(32) found that for each impact, the impact parameter correlated
well with flesh firmness by penetrometer (r = 0.86) and elastic modulus (r =
0.94). They also found that the same fruits can be subjected
to four multiple
impact tests without causing damage. Tissue damage after multiple drops at 40
mm height was evaluated to confirm that the test was nondestructive. They concluded that combination of multiple impact with a simple average gave better
results than using multiple regression. They also claimed that random multiple
impacts eliminate the need to specifically orient the fruit on a conveyor system
and is fast for real time sorting (0.2 s). However, it is only suitable for fruits that
do not bruise easily to ensure minimal damage during testing.
A common difficulty for all the firmness-measurement methods based on
nondestructive force is the need for mechanical contact between the product and
the sensor. Mechanical devices can limit the speed of operation and can limit
reliability. Speed is also limited in some of the approaches by the need to complete complicated analyses on the signal generated. There
is also a potential problem of cross-contamination from one product to another when touched by the
same probe.
Table 2 Firmness Measurement by Methods Based on Impact and Impact Rebound
h)
Princiole
Impact of fruit on
plate
Measurement
Force
Piezoelectric
force sensor
Product
Kiwifruit
Peach
Multiple impact of
fruit on surface
Impact rebound by
low-mass object
on fruit
Measurement condition
Performance of device
Store 4C
Test 20C
Drop height 10-50 mm
Test 23C
Drop height 50 mm
R? > 0.7 (device vs. Effegi penetrometer)
Research
65
r = 0.73-0.89 (device
vs. Effegi) cv =
0.057-0.062
R' = 0.47-0.83 (device
vs. penetrometer)
Reject 89 2 3% of soft
fruit (device)
Reject 96 2 1% of soft
fruit (human)
r = 0.86 (device vs.
Penetrometer)
r = 0.94 (device vs. parallel plate)
Accuracy = 7 8 4 9 %
(device vs. Instron)
Variability = 1.55%
(device)
S.D. = 12.9% (device)
S.D. = 15.3% (Instron)
R? = 0.78-0.84 (device
vs. MT)
No damage (20 g impactor)
32% damage for 59 g
impactor at 2 cm
Research
84
Research
33
Research
32
Research
85
Research
86
Research
87
Kiwifruit
Test room temp.
Cantaloupe
melons
Ambient
Drop height 2 crn
Peaches
Test room temp. (20"C,

85% RH)
Test 20C
Drop height 100 mm
Acceleration
(mass impactor)
Cherry
Force
Piezoelectric transducer
Pears
Store 0C
Test 2OoC, Drop height
2-4 cm
Impact mass 10-50g
Status
Ref.
h,
3
u)
2
'u
00
00
o \ D
!z
I
.
e,
3
E
Q
273
274
Hung et at.
Acoustic resonance
Sound velocity
Microphone
Melon
Test room temp. 2 mics, 6 cm apart
Store 2C
Test 20C
2 rnic. 1.5 cm apan
Kiwifruit Store 5C
Test 2 0 T
Ethylene-treated
Store 7 2 2C. 4 0 4 0 % RH
Peach
Test 20C I rnic
Store 1C. 96% RH
Apple
Test 20C
Tomato
Store and test 20C. I mic
Store 1C
Apples
Test 1C
I mic
Store 20C. 65% RH. Test 20C
Refleaion spectra IOOO-1650 nm
Apple
Vibration
Microphone
Vibration
Microphone
NIR spectroscopy
NMR (85.5 MHZ)
Microwave (0.2-20
GHz)
Laser doppler
Reflectance spectra (visible light

and PbS detectors)
Signal decay (relaxation)
Dielectric constant ( E ' ) and
loss factor (el')
Surface vibration
Doppler laser
Avocado
Test rmm temp
Tomato
Peach
Test r w m temp
Test 23 f I T
Kiwifruit Store 5C
Test 20C
Store 10C
Peach
Test 20C
Store 10C
Pea
Test 20C
r = 0.94 (transmission velocity vs. E)

r = 0.87 (transmission velocity vs. force)
r = 0.79 (transmission velocity vs. E)
r = 0.89 (transmission velocity vs. MT)
research
98
Research
87
Trend i n ultrasonic elapsed time agreed with

trend in penetrometer tests (60"probe)
Research 43
R'
Research 93
0 64-0.74 (device vs. Effigi)
n
-'
(D
u)
'p
u)
23
(D
R' = 0 84 (device vs. spherical probe)

R: = 0 84 (device vs. parallel plate)
R' = 0 7 6 (device vs E)
R2 = 0.27 (device vs MT)
r
r
=
=
0.75 (device vs. MT)

0 90 (device vs stiffness factor. f' m"')
Research 99
Research
100
u)
Research 59
Research
R:
Research
= 0 81-0.92 (device v 5 penetrometer. 30"

cone probe)
!2
= 0 98 (oil/water resonance peak ratio vs.

fruit dry weight)
R' = 0 85 ( I IT2 spin-spin time constant vs MT)
R' = 0.57-0.80 (device vs MT)
R'
101
Research 55
Research 102
53
Store = Stored at indicated temperature/condition; Test = temperature during test; Freq = frequency; mic = microphone; temp. = temperature: NMR =
nuclear magnetic resonance; PMR = proton magnetic resonance; NIR = near-infrared
h)
-I
vl
276
Hung et al.
C. MethodsBasedonSecondaryProperties
A number of methods basedon indirect measurement of fruit firmness have been
developed and investigatedby several researchers. These methods, most of which
are experimental, are based on the secondary properties of fruits, which could
be calibrated with a direct firmness-sensing method such as the
MT test.
Consumer studies by Hung et al. (19) showed that an experts ranking correlated well with the consumer panel ranking (r= 0.92) and that firmness among
all five firmness groups were significantly different. Sorting accuracies
of the
experts ranking comparedto the consumer results obtainedby discriminate analysis were 76, 73, 74, 66, and 87% for most firm, very firm, firm, less firm, and
least firm, respectively. This indicates that although human firmness sorting may
be subjective, it is still a reliable method.
Abbott (75) found that sonic measurements of Delicious apples correlated
not as
significantly with other destructive measurements but correlations were
good as the ripeness scores obtained
by the USDA inspectors. Mechanical impulse excitation and piezoelectric sensors were used
by Galili et al. (52)to nondestructively measure the firmnessof avocado fruits. They foundthat the correlation
coefficient, r, betweentheacousticparametersandthemaximumpenetration
force ranged from 0.66 to 0.8 1.
Kiwifruit, peach, and pea firmness was investigated by Muramatsu
et al.
(53, using a laser doppler vibrometer(LDV), with which there was no mechanical
contact with the product. They found that the phase shift between the received and
emitted wave by the fruit gave good correlations (R 5 0.9) with fruit firmness.
of deMuramatsu et al. (53) also reported that the LDV technique was capable
tecting citrus fruits afflicted with internal disorders.
Using NMR, fairly good correlations(r = 0.7 I ) were obtained by Stroshine
et al. (76) for cherries between elasticity and T? time as measured in a magnetic
resonance process analyzer. Kim et al. ( 5 5 ) found the signal decay rate from the
magnetic resonance was highly correlated with the firmness
of tomatoes. Laser
vision imagery employing an He-Ne laser at 632.8 nm wavelength and two diode
laser modules at 670 and 685 nm as light sources were used by Lee et al. (77)
to nondestructively measure the firmness of Golden Delicious, Fuji, and Rome
apples. They compared the firmness values with those obtained destructively by
the MT method(withoutpeeling theappleskin).Factorsthatinfluencedthe
firmness measurements included light type (monochromeor color), skin removal,
laser optical power, and laser wavelength.
The performance data for these and
other instrumental firmness measuring methods are summarized in Table 3.
D. NondestructiveMethodswith No MechanicalContact
Several direct comparisons of the laser air-puff deformations with MT measurements have been made for peaches yielding R? values of 0.79 (78), 0.75 (7), and
Methods
277
Table 4 Sorting Performance of Laser Air-Puff Detector and Magness-Taylor Test
Comoared to Packinghouse Manager and USDA Inspector

Magness-Taylor
detector air-puff
Packinghouse
manager
Laser
Predicted
soft
Predicted
firm
Confirmed 21%
79%
(miss)
soft
(hit)
Confirmed (correct
73%
27%
firmacceptance)
(false
(false firmacceptance)
alarm)
Magness-Taylor
detector air-puff
ceptance)
(false
Packinghouse Predicted
soft
manager
Confirmed
soft
Confirmed
Predicted
firm
75%
(hit)
25% (miss)
15%
85% (correct
alarm)
Laser
Predicted
firm
USDA
inspector
Predicted
soft
Confirmed
soft
Confirmed
firm
100%
0% (miss)
(hit)
(hit)
37% &
63% (correct
alaI?Tl)
USDA
inspector
Confirmed
soft
Confirmed
Predicted
soft
Predicted
firm
100%
0% (miss)
26%
(false
alarm)
74% (correct
acceptance)
0.86 (19). Correlations between modulus of elasticity values calculated by the

ASAE Standard for spherical indenter and those calculated from similar equations
for the laser air-puff were even higher for both apples (R = 0.91) and peaches
(R? = 0.92). Correlations with rubber balls have ranged from R? = 0.94 to 0.97
(7).
Table 4 demonstrates results from the laser air-puff andMT methods conipared with measurements by a packinghouse manager and a USDA inspector
(78). When the packinghouse manager was considered as the standard, the laser
air-puff had a slightly smaller miss
rate (21%) than the MT (25%). However.
the false alarm rate for the laser air-puff (27%) was higher than the
MT (IS%).
When compared with the USDA inspector, both the laser air-puff and the MT
had a zero miss rate. Again, the laser air-puff false alarm rate was higher (37%)
than the MT (26%).
VI.
APPLICATIONS
Nondestructive textural quality evaluations are made over a wide range of applications. Produce growers manually
feel products as they matureto help determine
278
Hung et al.
harvest date (7). Workers who harvest products also use tactual information to
evaluate maturity. Sorters on packing lines manually remove items that are too
soft or too fitm. Government inspectors use tactual input to verify that grade
standards are satisfied. Buyers judge quality by feeling the product. Consumers
predict eating quality of products based on firmness when purchasing products.
Mouthfeel judged by the consumer is the final evaluation of firmness.
A primary application for a nondestructive firmness sensor is
to replace
subjective human evaluations with a quantitative measurement. Objective measurements of fruit firmness would enable precise descriptionsof the condition of
shipments as they move from the grower
to the final consumer. Firmness data
from a calibrated instrument fora sample of product from a shipment would help
resolve differences in opinion about the condition of the product.
A relatively small sample size
is collected from shipments for firmness
evaluations when a destructive test is used because all of the product examined
must be discarded after the test. When usinga nondestructive instrument a much
larger sample size can be measured and returned to the containers. The increase
in sample size provides a stronger inference for the actual condition
of the lot
(shipment).
Researchers will benefit in a similar way as the private sector from equipment for making nondestructive firmness measurementson samples from storage
of productare
studies.Whenusingdestructivemeasurements,largeamounts
A nonneeded so that small samples can be removed periodically for evaluation.
to be returned to the study.
destructive measurement allows the product tested
Periodically following changes in the firmness of each item in the study also has
the advantage of providing the opportunities of modeling changes to each item
rather than modeling average changes.
The ultimate goal of nondestructive firmness sensing is detection of firmness on fruit and vegetable packing lines. On-line firmness sorting of each item
flowing through a packinghouse would improve the consistency
of shipments
and reduce the amount of good product discarded. Removal of the few overripe
of decay spreading to good
(soft) items in a container reduces the possibility
items. The end result is higher economic returns not only for the shipper but for
each business in the marketing chain. Finally, consumers benefit from improved
firmness sorting by having product at the desired firmness
at the point of purchase
and when consumed.
VII. SUMMARYAND FUTURETRENDS

The traditional approach of evaluating a new nondestructive instrument with the
existing destructive instrument (Magness-Taylor) could be limiting the opportunity to identify one or more new approaches with the potential for improving
279
overall food distribution systems. Another approach would be to evaluate a new

instrument by using economic comparisons of two handling systems: (a) destructive measurements of firmness on a small sample to make decisions on the acceptance or rejection of a shipment compared with (b) nondestructively measuring
to mechanically remove
the firmness of each item before packing a shipment
items that are too soft or too firm.
Over the past two decades, a number of methods and devices have been
used by different researchers to investigate the firmness of foods. The methods
are often based on the product being tested, intended application (bench-top,
field,
on-line), and mode of measurement, Le., destructive or nondestructive measurement. There is also an enormous range of food materials (natural and processed
foods), each behaving differently under different conditions. This makes
study
the
of food firmness and specifically the development of standardized nomenclature
of terminology difficult because each product requires a specific procedure and
units of expression.
to be nonhomogeneous and
Indeed, most biological products are known
anisotropic. Abbot and Lu (63) conducted compression tests on three cultivars
of apple using an Instron Universal Testing machine and evaluated the effects
of ripeness, specimen orientation, and location within the apple on failure stress,
strain, and energy, and apparent modulus of elasticity (Youngs modulus). They
found that mechanical properties were significantly influenced
by specimen orientation, latitude (location from stemto calyx), and depth (from skinto core). Interactions among these factors were also found to be significant.
Another problem associated with the study of food firmness is that it often
requires a reference or standard material. Should such a standard exist, it should
be structureless (have no atoms), isotropic (properties are independent of direction), and ideal (does
notexist in reality).However, it shouldbepossibleto
develop a realistic (if not perfect) standard product that could be purchased
to
enable calibration of nondestructive firmness detectors such as done for other
instruments (color tiles for calibrating colorimeters). Ideally, firmness standards
would have shapes that are similar to the products tested. One application of a
standard for firmness is to make sure an instrument reads the same after a test
as it did before the test. A standard could also be used for evaluating the performance of a new instrument compared to an existing one. A commercial application is to calibrate instruments at both shipping and receiving points.In all applia predeterminedfirmnessthatcouldbe
cationsthestandardwouldhave
reproduced or measuredbyindependentmethodssuchasusingdestructive
methods.
In addition, a distinction should be made between surface firmness and
gross firmness. Surface firmness refers to the measurement made when firmness
is determined based on the analysis of an input signal to, and response of, essentially the cellsof the product closeto the surfaceof the food. Examplesof surface
Hung et al.
280
firmness include puncture and laser air-puff. On the other hand, gross firmness
is determined when the whole food product receives the input signal and responds
with a measurable output response. Examples of gross firmness would be results
obtained from acoustic, NMR, and NIR measurements.
in this section indicate continuing
The many recent dates for work cited
interest in new approaches for measuring product firmness. However, there are
difficulties with all the attempts described above in all three general approaches.
of the need to extrapolate
Destructive methods have the inherent disadvantage
to represent the firmness of the batch or shipment.
the results from a sample
Methods that apply a controlled force to the product often have limited capacity
due to speed restrictions from inertial responses of moving masses. Methods that
measure a secondary property suffer from difficulties
in relating the results to
the firmness property. The Washington Fruit Research Commission conducted
evaluations of several nondestructive firmness-measuring devices in 1994. The
devices included an impact system in which a rod strikes the fruit with the force
of the impact converted to a firmness reading, measuring sound resonance when
the fruit is tapped by a plastic hammer, and by ultrasound (79). The results of
all the devices had very weak relationships with readings from Magness-Taylor
tests. Thus, a need is evident for a new solution
to the problem of measuring
firmness. However, Warner (80) contended that the fruit industry is searching
for firmness-sorting devices that need not be compared with the Magness-Taylor
method, which has its limitations, but devices that are themselves capable of
demarcating soft from firm fruit on an absolute basis.
To improve the accuracy and dependability of on-line firmness measurements, a concept of sensor fusion, that is, combining two or more sensing methods
(8 1 ) attempted to automate fruit sorting
is currently being investigated. Ozer et al.
by fusion of data acquired from selected sensors. The information obtained was
used to classify cantaloupe fruits into four maturity stages. Also, Zhang et al.
(82) presented a method based on fuzzy modeling
of the overall quality of a
product based on a setof quality factors, F (which included firmness), and another
p. Because each individset of grading factors,G, to obtain a single grading factor
ual method for firmness evaluation hasits own limitations, sensor fusion can take
the advantage of several sensing methods based on different sensing principles
to obtain the best prediction/measurement of firmness. We believe that Sensor
fusion (Fig. 1 ) is the correct direction for future firmness sensing.
REFERENCES
I.
AC Mackey, MM Hard. MV Zaehringer. Measuring textural characteristics of fresh

fruitsandvegetables-apples,carrots.andcantaloupes.TechnicalBulletin
123,
Agricultural Experiment Station. Oregon State University, Corvallis.
OR, 1973.
ethods
2.
281
s Tsuji, Recent research in food texture (mouthfeel) and its instrumental measure-
ment in Japan. J Texture Stud 15:19.5-225. 1984.

3. A Kramer. Texture-its definition, measurement and relation to other attributes of
food quality. Food Technol 26( 1):34-39. 1972.
4. T Izutsu, K Wani. Food texture and taste: a review. J Texture Stud 16:1-28. 1985.
5. AS Szczesniak. Classification of textural characteristics. J Food Sci 28:385-389.
1963.
6. HJ Raeuber, H Nikolaus. Discussion paper-structure of foods. J Texture Stud 1 1:
187- 198, 1980.
7. y-c Hung, SE Prussia, GO1 Ezeike. Nondestructive maturity sensing using a laser
airpuff detector. Postharvest Biol Technol 16: 15-25, 1999.
8. WAC 16-403-142, Red Delicious. Delicious and Golden Delicious minimum Standards. Washington State Standards, Olympia, WA, 1990.
9. A Pedrosa-Menabrito, JM Regenstein. Shelf-life extension of fresh fish-A review
Part III-Fish quality and methods of assessment. J Food Qual 13:209-223, 1990.
10. MC Bourne. Food Texture and Viscosity: Concept and Measurement. New York:
Academic Press Inc, 1982: 44-1 17.
1 1 . J Giese. Measuring physical properties of foods. Food Technol 49(2):54-63, 1995.
12. p Chen, Z Sun. A review of nondestructive methods for quality evaluation and
sorting of agricultural products. J Agric Eng Res 49:85-98, 1991.
13. PM Voisey, E Larmond. Texture of baked beans-a comparison of several methods
of measurements. J Texture Stud 2( 1):96-109, 1971.
14. FR Harker, MGH Stec, IC Hallett, CL Bennett. Texture of parenchymatous plant
tissue: a comparison between tensile and other instrumental and sensory measurements of tissue strength and juiciness. Postharvest Biol Technol 11:63-72, 1997.
15. GG Dull. Nondestructive quality evaluation of agricultural products: a definition
and practical approach. J Food Prot 41(1):50-53, 1978.
16. FR Magness, GF Taylor. An Improved Type of Pressure Tester for Determination
of Fruit Maturity. USDA Circular 350, Washington. D.C., 1925.
17 CJ Studman, Yuwana. Twist test for measuring fruit firmness. J Texture Stud 23:
215-227,1992.
18. J A Murillo, GH Brusewitz, NO Maness. Peach texture during ripening after extended storage. J Food Qual 205-72, 1997.
19. Y-C Hung, KH McWatters, SE Prussia. Sorting performance of
a nondestructive
laser air-puff firmness detector. Appl Eng Agric 14(5):513-516, 1998.
20. JJ Mehlschau, P Chen, LL Claypool. RB Fridley. A deformeter for nondestructive
maturity detection of pears. Trans ASAE 24(5):137 I - 1375, 1981.
21. EJ Timm, GK Brown, PR Armstrong,RM Beaudry. A portable instrument for
Incasuring firmness of cherries and berries. Paper
No. 93-6.539. St. Joseph. MI: ASAE.
December 14- 17, 1993.
22. A Mizrach, D Nahir. B Ronen. Mechanical thumb sensor
for fruit and vegetable
sorting. Trans ASAE 3.5247-250, 1992.
23. H Takao, S Ohmori. Development of device for nondestructive evaluation of fruit
firmness. JARQ 28( l):36-43, 1994.
24. JR Botta. Instrument for nondcstructive tcxture measurement of raw Atlantic Cod
(Gnd~tsrrlorhrtr) fillets. J Food Sci 56(4):962-968. 1991,
282
Hung et al.
25. IJ Davie, NH Banks, PB Jeffery, CJ Studman, P Kay, P. Nondestructive measurement of kiwifruit firmness. NZ J Crop Hort Sci 24:151- 157, 1996.
26. GA Dawson. Non-destructive firmness testing of kiwifruit. Programme information
and abstracts of the second International Symposium on Kiwifruit, Massey University, Palmerston North, New Zealand, 18-21 February, 1991.
27. PR Armstrong, GK Brown, EJ Timm. Non-destructive firmness measurement of
soft fruit for comparative studies and quality control. Paper No. 95-6172,
St. Joseph,
MI:ASAE,June18-23,1995.
28. JA Abbott, DR Massie. Nondestructive firmness measurement of apples. Paper No.
93-6025, ASAEKSAE meeting presentation. St. Joseph, MI: ASAE, 1993.
29. MJ Delwiche, T McDonald,SV Bowers. Determinationof peach firmnessby analysis of impact forces. Trans ASAE 30(1):249-254, 1987.
30. MJ Lichtensteiger, RG Holmes, MY Hamdy, JL Blaisdell. Impact parameters of
spherical viscoelestic objects and tomatoes. Trans ASAE 31:595-602, 1988.
31. MP Rigney, GH Brusewitz, ML Stone. Peach physical characteristics for orientation. Trans ASAE 39(4): 1493-1497, 1996.
32. N Ozer,BA Engel, JE Simon. A multi impact approach for nondestructive measurement of fruit firmness and maturity. Trans ASAE 41(3):871-876, 1998.
33. PN Schaare, VAMcGlone.Designandperformanceof
a fruitfirmnessgrader.
Proceedings of Sensors for nondestructive testing-measuring the quality of fresh
fruits and vegetables. Orlando, Northeast Regional Agricultural Engineering Service, Ithaca, NY. Pp 27-36, 1997.
34. RW Bajema,GMHyde.Instrumentedpendulumforimpactcharacterizationof
whole fruit and vegetable specimens. Trans ASAE 41(5):1399-1405, 1998.
35. P Chen, M Ruiz-Altisent, P Barreiro. Effect of impacting mass on firmness sensing
of fruits. Trans ASAE 39(3):1019-1023, 1996.
36. WL Bryan, GJ Anderson, JM Miller. Mechanically assisted grading of oranges for
processing. Trans ASAE 21:1226- 1231, 1978.
37. R Feller, D Nahir, CG Coble. Separation of soil clods from onions using impact.
Trans ASAE 27:353-357, 1984.
38. R Feller, E Margolin, A Zacharin, H Pasternak. Development of a clod separator
for potato packing houses. Trans ASAE 28:1019-1023. 1985.
39. DR Bower, RP Rohrbach. Application of vibrational sorting to blueberry firmness
separation. Trans ASAE l9(1):185-19l. 1976.
I):
40. K Peleg. Development ofa commercial fruit firmness sorter.J Agric Eng Res 72(
23 1-238, 1999.
41. JS Perry. A nondestructive firmness testing unit for fruit. Trans ASAE 20:762767,1977.
42. CN Thai. Thickness effects of soft transducer used for firmness measurement. Paper
No. 95-61 IO. ASAE, St. Joseph, MI: 1995.
43. N Muramatsu,N Sakurai, R Yamamoto,DJ Nevins, T Takahara, T Ogata. Comparison of nondestructive acoustic method with an intrusive method for firmness measurement of kiwifruit. Postharv Biol Techno1 12:221-228, 1997.
44. J Sugiyama, T Katsurai,J Hong. H Koyama, K Mikuriya. Melon ripeness monitoring by a portable firmness tester. Trans ASAE 41( 1): 12I - 127, 1998.
45. EE Finney. Mechanical resonance within Red Delicious apples and its relation to
fruit texture. Trans ASAE 13(2):177-180, 1970.
Methods
283
46. EE Finney. Random vibration techniques for non-destructive evaluation of peach

firmness. J Agric Eng Res 16(1):81-87, 1971.
47. EE Finney. Vibrational techniques for testing fruit firmness. J Texture Stud 3(3):
263-283,1972.
48. JA Abbott, LA Liljedahl. Relationship of sonic resonant frequency to compression
testsandMagness-Taylorfirmness of applesduringrefrigeratedstorage.Trans
ASAE 37(4):121 1-1215, 1994.
49. NGalili,AMizrach, G Rosenhouse.Ultrasonictesting ofwholefruit fornondestructivequalityevaluation.PaperNo.93-6026.StJoseph,MI:ASAE,
1993.
50. A Mizrach,U Flitsanov, Y Fuchs. An ultrasonic nondestructive method for measuring maturity of mango fruit. Trans ASAE 40(4): 1107-1 11 1, 1997.
51. Y Cheng. Non-destructive quality evaluation of fruits and vegetables using ultrasound.Ph.D.dissertation,VirginiaPolytechnicInstituteandStateUniversity,
Blacksburg,VA,1992.
52. N Galili, I Shmulevich, N Benichou. Acoustic testing of avocadofor fruit ripeness
evaluation. Trans ASAE 41(2):399-407, 1998.
53. N Muramatsu, N Sakurai, N Wada, R Yamamoto, T Takahara, T Ogata,K Tanaka,
T Asakura, Y Ishikawa-Takano, DJ Nevins. Evaluation of fruit tissue texture and
internal disorders by laser doppler detection. Postharvest Biol Technol 15:83-88,
1999.
54. N Muramatsu, N Sakurai, N Wada, R Yamamoto, T Takahara, T Ogata, K Tanaka,
TAsakura, Y Ishikawa-Takano, DJ Nevins.Remotesensingoffruittextural
changeswith a laserdopplervibrometer. J AmSOC Hort Sci125(1):120-127,
1999.
5 5 . S Kim, MJ McCarthy, P Chen. Feasibility of tomato quality grading and sorting
using magnetic resonance. Paper No. 94-6519. St Joseph, MI: ASAE, 1994.
56. CJ Clark, PD Hockings, DC Joyce, RA Mazucco. Application of magnetic resonance imaging to pre- and post-harvest studies of fruits and vegetables. Postharvest
Biol Technol 11: 1-21, 1997.
57. S Gunasekaran. Delayed light emission as a means of quality evaluation of fruits
and vegetables. Crit Rev Food Sci Nutr 29(1):19-34, 1990.
58. Y Chuma, K Nakaji. Delayed light emission as a means of color sorting of plant
products. In:P Linko, Y Malkki, J Olkku, eds. Food Process Engineering, Vol 1.
London: Applied Science Publishers, 1979:3 14.
59. J Lammertyn, B Nicolai. K Ooms, V De Smedt, J De Baerdemaeker. Nondestructive measurement of acidity, soluble solids and firmness of JonaGold apples using
NIR-spectroscopy. Trans ASAE 41(4): 1089-1094, 1998.
60. P Jeffery, NH Banks. Firmness-temperature coefficient of kiwifruit. NZ
J Crop
Hort Sci 22:97-101, 1994.
61. RL Jackman, DW Stanley. Failure mechanisms of tomato pericarp tissue suggested
by large and small deformation tests. J Text Stud 23:475-489, 1992.
62. GO1 Ezeike. L Otten. Two-compartment model for drying Egusi (melon) seeds.
Can Agricu Eng 33( 1):73-78, 1991.
63. JA Abbott, R Lu. Anisotropic mechanical properties of apples. Trans ASAE 39(4):
1451-1459,1996.
64 NO Maness,GH Brusewitz. Performance of an instrument designed for, and evalua-
284
65.
66.
67.
68.
69.
70.
71
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
Hung et al.
tion of methods to assess, peach impact bruise susceptibility. J Food Text 18:335353,1995.
VA McGlone, RB Jordan, PN Schaare. Mass and drop height influence on Kiwifruit
firmness. Trans ASAE 40(5):1421-1428, 1997.
HChen, DJ Baerdemaeker.Optimizationofimpactparametersforreliableexcitationofapplesduringfirmnessmonitoring.
J AgricEngRes
61:275-282,
1995.
SE Prussia, JJ Astleford, Y-C Hung, R Hewlett. Non-destructive firmness measuring device. U S . Patent 5,372,030 (December 13, 1994).
S Fan. Engineering approaches for improving the
UGA laser air-puff food firmness
detector. M.S. thesis. The University of Georgia, Griffin, Georgia, 1995.
VA McGlone, RB Jordan, PN Schaare. Anomalous firmness changes in cool-stored
kiwifruit. Postharvest Biol Technol 12: 147-156, 1997.
ASAE.Compressiontestoffoodmaterialsofconvexshape.StandardS368.3,
ASAE Standards. 44th ed. St. Joseph, MI: American Society of Agricultural Engineers,1997.
SE Prussia, GO1 Ezeike,Y-CHung.Performanceevaluationsofnondestructive
techniques for measuring produce quality (abstr). Sensors for Nondestructive Testing International Conference, Orlando, FL, 1997.
SE Prussia. Dynamic visual inspection. In: Human factors teaching modules. DL
Roberts, WI Becker eds. St. Joseph, MI: ASAE, 1991.
JA Abbott, AE Watada, DR Massie. Effegi, Magness-Taylor, and Instron fruit pressure testing devices for apples, peaches, and nectarines.J Am Soc Hort Sci lOl(6):
698-700,1976.
YN Wan, FH Buelow,S Gunasekaran. Engineering firmness study
of fruit maturity.
Paper no 92-6513. St. Joseph, MI: ASAE, 1992.
JA Abbott. Firmness measurement of freshly harvested Delicious apples by sensory methods, sonic, transmission, magness-Taylor, and compression.
J Am Soc
HortSci119(3):510-515,1994.
RL Stroshine, SL Cho, WK Wai, GW Kurtz. Magnetic resonance sensing of fruit
firmness and ripeness. ASAE Paper No. 91-6565. St. Joseph, MI: ASAE, 1991.
J-W Lee, YJ Han, Y-J Cho. Nondestructive firmness measurement of apples using
laser vision system. ASAE Paper No. 97-3081. St. Joseph, MI: ASAE, 1997.
Y-CHung, SE Prussia. Firmness measurement using a nondestructive laser airpuffdetector.ProceedingsoftheFoodProcessingAutomationConferenceIV,
McCormick Place, Chicago, IL, 1995.
SL Sanger. Pressure tester searchhas changed. Good Fruit Grower 45(16):7, 1994.
G Warner.Automaticsortingandsizingarehere,
Is firmnessnext?GoodFruit
Grower 44(6):18- 19, 1993.
N Ozer, BA Engel, JE Simon. Fusion classification techniques for fruit maturity.
Trans ASAE 38(6):1927-1934, 1995.
Q Zhang, A Breen, JB Litchfield. A fuzzy comprehensive evaluation model for
food products grading. Paper No. 89-6613. St. Joseph, MI: ASAE, 1989.
P Lesage, M Destain. Measurement of tomato firmnessby using a nondestructive
mechanical sensor. Postharvest Biol Technol 8:45-55, 1996.
X Zhang, ML Stone, D Chen, NO Maness, GH Brusewitz. Peach firmness determi-
285
nation by puncture resistance, drop impact, and sonic impulse. Trans ASAE 37(2):
495-500,1994.
for nondestructive deter85. FI Meredith, RG Leffler, CE Lyon. Impact force response
mination of peach firmness. Proceedings of the Food Processing Automation conference, Lexington, KY: 1990, pp 212-217.
86. EJ Mitcham, M Clayton, WV Biasi. Comparison of devices for measuring cherry
fruit firmness. HortScience 33(4):723-727, 1998.
87. H Chen, F Duprat, M Grotte, D Loonis, E Pietri. Relationship of impact transmission wave to apple texture during ripening. J Text Stud 27: 123-141, 1996.
88. MJ Delwiche, H Arevalo, J Mehlschau. Second generation impact force response.
Trans ASAE 39(3):1025-1033, 1996.
89. M Delwiche,Y Sarig. A probe impact sensor for fruit firmness measurement. Trans
ASAE 34(1):187-192, 1991.
90. MJ Delwiche, H Arevalo, J Mehlschau. A fruit firmness sorting system. Paper NO.
93-6024, St Joseph, MI: ASAE, 1993.
91. FL Younce, DC Davis. A dynamic sensor for cherry firmness. Trans ASAE 38(5):
1467-1476,1995.
92. ML Stone, PR Armstrong, DD Chen, GH Brusewitz, NO Maness. Peach firmness
prediction by multiple location impulse testing. Trans ASAE 41(1): 115-1 19, 1998.
93. PR Armstrong, ML Stone, GH Brusewitz. Peach firmness determination using two
different nondestructive vibrational sensing instruments. Trans ASAE 40(3):699703,1997.
94. RL Clark, PS Shackelford. Resonance and optical properties of peaches as related
to flesh firmness. In: JJ Gaffney, ed. Quality Detection in Foods. St Joseph, MI:
ASAE,1976:143-145.
95. JA Abbott, HA Affeldt, LA Liljedahl. Firmness measurement of stored Delicious
apples by sensory methods, Magness-Taylor, and sonic transmission.
J Am Soc
HortSci117(4):590-595,1992.
96. JA Abbott, DR Massie, BL Upchurch, WR Hruschka. Nondestructive sonic firmness measurement of apples. Trans ASAE 38(5): 1461 -1466. 1995.
97 A Mizrach, U Flitsanov. Nondestructive ultrasonic determination of avocado softening process. J Food Eng 40: 139- 144, 1999.
98. J Sugiyama, T Katsurai, J Hong, H Koyama, K Mikuriya. Melon ripeness monitoring by a portable firmness tester. Trans ASAE 41(1 ): 121- 127, 1998.
99. F Duprat, M Grotte, E Pietri, D Loonis. The acoustic impulse response method for
measuring the overall firmness of fruit. J Agric Eng Res 66:251-259, 1997.
100. P Armstrong. HR Zapp, GK Brown. Impulsive excitation of acoustic vibrations in
apples for firmness determination. Trans ASAE 33(4): 1353- 1359, 1990.
101. PChen, MJMcCarthy,S-MKim,
BZion.Development of ahighspeed NMR
technique for sensing maturity of avocados. Trans ASAE 39(6):2205-2209, 1996.
102. SO Nelson, WR Forbus Jr, KC Lawrence. Assessment of microwave permittivity
for sensing peach maturity. Trans ASAE 38(2):579-585, 1995.
Linear Viscoelastic Methods

M. Mehmet Ak
Istanbul Technical University, Maslak, Istanbul, Turkey
1.
INTRODUCTION
Food quality evaluation by consumers is a dynamic and complex process, which

varies from person to person and also depends on the food and situation (1 -3).
Nevertheless, food quality has long
been investigated by sensory and instrumental
means, often to establish a correlation between results from the two methods
(43). Since food quality is a multidimensional and complicated subject,
it is
generally studied in the context of subcategories such as appearance, flavor, texture, nutrition, safety, convenience,and stability (3,4). Hence, instrumental quality evaluation within a subcategory (e.g., texture) can
be expected to provide
useful but partial information towards better understanding
of the overall food
quality ( 1 -3).
Historical developments making texture one of the key attributes of food
quality have been discussedin a recent articleby Szczesniak, one of the pioneers
in the field of food texture(6). The influenceof textural attributeson the consumers perception of food quality changes depending upon the product (4,6,7). The
textural attributesof foods are usually estimated
by trained sensory panelists and/
or by a variety of instrumental methods (4,5,8-10). During sensory analysis of
food texture, samples are often disintegrated (e.g.,
by mastication or pressing)
in order to perceive textural attributes. Accordingly, the
most widely used instrumental texture evaluation method [i.e., texture profile analysis (TPA)]
is devised
to simulate the action of the jaw (4). Therefore, it is also a destructive test. In a
287
288
Ak and Gunasekaran
typical TPA test a sample is subjected to two consecutive compressions (double

bites) that are usually beyond the point
of fracture or gross failure of foods.
to determine macroscopic textural attributes
of
Such studies have been useful
foods. Food texture evaluation
by destructive methods (e.g., TPA) and correlation
research
of instrumental data with sensory results continues to be an activeof area
(11).
Destructive techniques are often not sufficiently sensitive to elucidate the
structure-property relationships in food materials. For instance, physicochemical
reactions taking place in fruits and vegetables during maturation lead to several
as an indicator of
changes, including softening of the tissue, and can be used
product quality (ripeness). Puncture tests or uniaxial compression tests are traditionallyused to evaluate fruit softening.Theselarge-deformation,destructive
tests provide data related to failure properties rather than microstructural changes
at the cellular level. Such large deformation tests to assess small-scale changes
might be considered inappropriate, if only because of their relative insensitivity
(12). However, small deformation tests, generally involving strains of less than
3%, permit evaluation of microstructural changes at the cellular level and facilitate characterizing time-dependent or viscoelastic properties thatall plant materials exhibit ( 13- 15).
Accordingly, much of recent research has been directed towards understanding of structure-property relationships in foods. In these studies, the aim
is to measure properties during and after structure formation without perturbing the process. Thus, the introduction of dynamic oscillatory method has been
animportantstep,whichenabledsimultaneousmonitoring
of viscousand
elastic responses of foods in a nondestructive manner. Ross-Murphy (16) pointed
out that NMR spectroscopy enables structural informationof foods to be obtained
at the molecular level (over distances of 0.5-5 nm), whereas mechanical spectroscopy enables foods to be probed over supramolecular distances ( 1 pm to 10
mm).
Rapid improvements in capability of rheological instruments, measurement
and analysis methods, and continuous need to control and/or modify material
in an impressive amount of
properties for high-quality products have resulted
publications on viscoelastic properties of foods. These studies have been very
useful in revealing the microscopic textural attributes of foods.
This chapter is comprised of two main sections. The first section is about
viscoelastic behavior, where the general characteristics of viscoelasticity, the utility of mechanical models in describing viscoelastic behavior in transient and dynamic tests, and the features of small amplitude oscillatory shear (SAOS) tests
are presented. In the second part, applications of the transient and dynamic tests
are presented. Here, the examples are selected to show the diverse utility of the
linear viscoelastic tests, particularly of SAOS,in studying viscoelastic properties
affecting quality of different food systems.
289
II.
VISCOELASTICITY
A.
IdealElasticandViscousBehaviors
An ideal solid material will respond to an applied load by deformingfinitely and

recovering that deformation upon removal of the load. Such a response is called
a direct
elastic. Ideal elastic materials obey Hookes law, which describes
and strain (y) via a proportionality constant
proportionality between the stress(0)
called modulus (G), i.e.,
0=
Gy
It should be noted thatG is more accurately called shear modulus (or modulus of rigidity) because it is obtained via a shear test. However, similar relationa symbol E is used to
ships hold good for tension and compression tests, and
represent Youngs modulus obtained via these tests. Since shear
is the favored
in this
mode of deformation in linear viscoelastic studies, we limit our discussion
chapter only to shear tests unless stated otherwise. For a perfectly elastic solid
E and G are related as E = 3G (1 7).
An ideal fluid will deform and continue to deform as long as the load is
applied. The material will not recover from its deformation when the load
is
removed. This responseis called viscous. Theflow of simple viscous materials
is described by Newtons law, which constitutesa direct proportionality between
i.e.,
the shear stress and the shear rate
(v),
The proportionality constant q is called the shear viscosity. From energy

of energy expended
considerations, elastic behavior represents complete recovery
during deformation, whereas viscous flow represents complete loss of energy as
all the energy supplied during deformation is eventually dissipated as heat.
B. TimeFactor in Viscoelasticity
An important feature of viscoelastic materials, in contrast to the ideal materials,
is that their response depends on the time-scale of observation. A viscoelastic
material may behave more like viscous liquid with some elastic effects, or like
on how it is studied. Generally,
elastic solid with some viscous effects, depending
the faster the deformation, the closer the response
is to being elastic, and the
slower the deformation, the closer the response
is to being viscous. Such behavior
results from the molecular structure
of the material: stretching intermolecular
bonds, which causes an elastic response, is very quick; the motion of molecules
past one another, which causes a viscous response, takes much more time ( 1 8).
The time scaleof experiments should be considered
in relation to the characteristic time, denoted by h, of materials (e.g., relaxation time).The character-
Ak and Gunasekaran
290
istic time of the materials varies widely-infinite for ideal elastic solids and zero
for ideal viscous liquids. A dimensionless number, Deborah number (De), is defined to relate h with the time-scale of observation (t) (19,20).
D, = hlt
A high Deborah number (De >> 1) corresponds to solid-like behavior (i.e., relaxation not observed), whereas a low Deborah number (D, << I ) corresponds to
a liquid-like response (i.e., fast relaxation observed). The viscoelastic behavior
(i.e., finite relaxation observed) will dominate when De is on the order of I .
From the definition of De, a material can exhibit solid-like behavior due
either to the long characteristic time orto rapid deformation processes (e.g., high10
speed testing). For instance, water has practically zero relaxation time,
s, and consequently is perceived as liquidin most cases (20). However, water can
respond as a solid material if the time of testingis very short (or equivalentlyif
the frequency of testingis very high). During production, handling, and consumption, food materials are subjected
to a wide variety of deformation processes,
in Fig. 1, the range of shear
each with a different time scale. As can be seen
rates, or time scales, in various processes (not all food-related) differs greatly.
In general,however,foodsarecomplexsystemsand
do notpossess a single
relaxation time but rather a broad range of relaxation times (21). Therefore, in
a given process of finite duration, foods are more likelyto behave as viscoelastic
materials. The significance of viscoelastic behavior and associated rheological
properties in extrusion processes, for example, is discussed in several articles in
Kokini et al. (22).
At this stage,it is interesting to mention that rheologists have recently been
deprived of a delightful and mythical example often usedto emphasize the effect
of the time scale of observation on the material response.
To illustrate liquidto an observation
like behavior of a solid material, rheologists sometimes referred
that windowpanes of old cathedrals are thicker at the bottom than at the top.
The
explanation was that the glasses of these ancient windowpanes slowly flowed at
ambient temperatures under the influence of gravity. Given that the time scale
as a plausible explanation
of flow was several hundred years, this was regarded
(23,24). However, earlier, Ernsberger (25) pointed out that at ordinary temperatures, glasses of commercially useful compositions are rigid solids and observation of thicker ancient windowpanes cannot be related
to flow due to gravity.
Plumb (26) pointed out that the correct explanation for thicker windowpanes at
the bottom liesin the glass processingof the ancient times. Recently, the explanation based on sagging of glass due to gravity was disproved by Zanotto (27),
who calculated that it would require more than the age
of the universe for a
typical window glass to flow appreciably so that the bottom would be thicker.
Viscoelastic materials are characterized by various phenomena observable
under different conditions. Some of these are (28):
astic
Linear
291
TIME (s)
SHEAR RATE (SI)
mm
m
Fig. 1 Shear rate ranges (and time scale of operation) for various food and industrial
processes. A: Sedimentation of fine particles in a suspending liquid (spices in salad dressing, medicines, paints); B: leveling due to surface tension (frosting, paints, printing inks);
C: draining under gravity (vats, small food containers, painting, and coating);
D: extrusion
(snackandpetfoods,cereals,pasta,polymers);
E: calendaring(doughsheeting);
F:
chewing and swallowing (foods); G: dip coating (paints, confectionery); H: mixing and
stirring (food processing); J: pipeflow(foodprocessing,blood
flow); K: spraying and
brushing (spray-drying, painting, fuel atomization);L: rubbing (applicationof creams and
lotions to theskin); M: high-speedcoating(paper); N: lubrication(bearings,gasoline
engines.)
If the stress is held constant, the strain increases with time (creep).
If the strain is held constant, the stress decreases with time (relaxation).
Effective stiffness depends on the rate of application of the load.
If cyclic loading is applied, hysteresis (or phase lag) occurs, leading to
dissipation of mechanical energy.
Acoustic waves experience attenuation.
Rebound of an object following an impact is less than 100%.
Frictional resistance occurs during rolling.
Ak and Gunasekaran
292
The viscoelastic response in a single physical entity can

be manifest in
many ways, some of which are conceptualized in terms of viscoelastic functions.
A thorough understanding of viscoelasticity entails a simultaneous appreciation
of all these manifestations and their analytical representations. Each viscoelastic
function, if known over the full range of time or frequency, contains complete
information regarding the linear behavior of a material. Nevertheless, each viscoelastic function emphasizes a different aspect of the behavior.
It is interesting to note that viscoelasticity
is the basis for the informal
nondestructive testing of ripeness or maturity of some foods by tapping. For
to judge
centuries, people have been tapping foods like melons and coconuts
their ripeness from the sound. This
is an example of free decay of vibration
following an impulse, a viscoelastic phenomenon (28).
C.
Viscoelastic Behavior in Transient Tests
Ideal elastic and ideal viscous behaviors present two extreme responses
of materials to external stresses. As the terms imply, these are only applicable for ideal
materials. Real materials, however, exhibit a wide array
of responses between
ideal viscous and elastic behaviors. Most materials exhibit both viscous and elastic properties simultaneously and, therefore, are characterizedas viscoelastic. Almost all foods belong to this group.
Ideal elastic and ideal viscous behavior is represented
by a spring and a
2). Thus,a
dashpot(a fluid-filled, piston-cylindersystem),respectively(Fig.
spring by itself can adequately represent a Hookean deformation, and the dashpot,
the Newtonian flow. Since viscoelastic materials have both viscous and elastic
properties, they are simply represented by a combination of spring(s) and dashpot(s).
The series combination of a spring and a dashpot
is called a Maxwell model
(Fig. 2). In this, both the spring and dashpot experience the same total stress ((3)
but share the total strain (y).
(3
ch=
(3d
(4)
y = 1: + yd
The subscripts s and d represent spring and dashpot, respectively. From the above,
the following Maxwell equation is obtained:
where, T = q,,/G.; and qd = q. One of the inadequacies of the Maxwell model

is that, for a constant stress (i.e., d d d t = 0) it reduces to Newtonian behavior,
which is generally not observed with viscoelastic materials in a creep test. How-
Methods Linear Viscoelastic
293
Output
nseFunction
Input Model
Mechanical
Element@)
Spring
Hookean
YC
Newtonian
Dashpot
Maxwell
Relaxation
Stress
u = ooexp(-fflR)
:Lh_
Kelvin
Y =ro[l-ex~(-ff~~)l
Fig. 2 Some simple mechanical elements used to model viscoelastic behavior of foods
and a set of their input-output relations.
ever, the Maxwell model is particularly useful in stress relaxation experiments,

for which it yields:
where, T~ is calledtherelaxationtime.Stressrelaxationresultsareoftenexpressed in terms of a material function called the shear relaxation modulus, G(t):
Gunasekaran
294
Ak and
The Maxwell model with a single exponential term is generally not sufficient to describe relaxation behavior of foods. Therefore, it is often modified to
include multiple exponential terms, each with a different relaxation time, representing an array of Maxwell elements in parallel (29). As can be seen in Eq. (7),
the Maxwell model predicts a zero stress at very long times>>
(t T ~ ) .That means
the model cannot represent a truly viscoelastic solid, which would have residual
stress after relaxation, as observed in many food products (30). To account for
the residual stress, the Maxwell model is modified
by adding a constant term
(31).
The parallel combination of a spring and a dashpot is called a Kelvin (or
Voigt or Kelvin-Voigt) model (Fig. 2).In this, both the spring and dashpot experience the same total strain (y) but share the applied stress (0).
d =
os+ dd
=
yd
(9)
(10)
From the above, the following Kelvin equation is obtained:

d =
Gy
+ qj
(1 1)
where G = G, and qd = q. One of the inadequacies of the Kelvin model is that,

= 0) it reduces to Hookean behavior, implying
a
for a constant strain (dyldt
constant stress at a constant strain, which is generally not observed with viscoelastic materialsin a stress relaxationtest. However, theKelvin model is of particular value in creep experiments, for which it yields:
where T~ is the retardation time. Creep results are often expressed in terms of a
material function called the shear creep compliance, J(t):
The Kelvin model can be generalized by combining many elements in series, each with a different retardation time (29). As can be seen in Eq. (12), the
Kelvin model predicts an asymptotically reached strain (y(J at very long times (t
>> 7,). That means the model cannot represent viscoelastic liquids showing continual flow. A viscous term is then included in the Kelvin model to account for
flow at long times (32).
The Maxwell and Kelvin elements are the simplest
of many mechanical
models. Many combinationsof several springs and dashpots (and other elements)
are used to fully describe the viscoelastic nature ofreal materials (33). For exam-
295
ple, the Burgers model, another common model used

to represent rheological
behavior of food materials, is a series combination of the Maxwell and Kelvin
elements.
The advantage of the mechanical models is that the elements of a model
can be related to individual parts of the structural network of the material being
studied. Fundamental rheological parameters obtained from mechanical models
provide information about the dependence of food structure on composition, interaction among constituents, and processes(34). For example, Dienerand Heldman (35) used a viscous Maxwell-Bingham rheological model to evaluate the
mechanical structure of butter.
The viscoelastic parameters relaxation time, zR,and retardation time, zD,
are experimentally determined,and the viscous and elastic parametersq,, and G,,
respectively, are computed (7, = zD = qd/G,). The relaxation modulus, G(t), and
5 l/J(t). The limiting case of G =
creep compliance, J(t), are related as G(t)
1 /J is applicable for a solid material following Hooke's law (17,18).
Because the time response of viscoelastic materials (y vs. t in creep or o
vs. t in relaxation) is nonlinear, it is difficult to discern the linear range from
experimental data at one strain (or stress) level. In order to determine the linear
range, then, multiple experiments must be performed at different constant strain
(or stress) levels. The linear range of the isochronal-the plot of stress against
strain at a specific time (in case of stress re1axation)"will indicate the extent
of strain level over which the material response can be considered linear. The
process of obtaining isochronal plots is illustrated in Fig. 3. The data obtained
(y, and y2) arerepresented in Fig. 3A andB.From
at twoconstantstrains
Fig. 3 Determininglinearviscoelasticrange from a set of stressrelaxationdata. (A)

Data obtained at an applied strain of y,; (B) data obtained at y2; (C) isochronals plotted
using data (a, b, c, and d) from A and B at times t, and tz.
Gunasekaran
296
Ak and
these, data points (for, e.g., a, b, c, d in Fig. 3) are gathered at different times
o versus y plot is constructedfor
(e.g., t,and t2). Thenthecorresponding
each of the times at which the stress response is measured (Fig. 3C). The strain
by dotted
value at which the isochronal beginsto deviate from linearity (indicated
line in Fig. 3C) is the upper limitof the linear viscoelastic region for the material.
Thehatchedregion in Fig.3Cindicatesthenonlinearrangeofthematerial
studied.
Performing experiments within the linear range is helpful because the data
can be analyzed using relatively simple linear viscoelastic models discussed previously. Within the linear range, the material functions (modulus and compliance)
are independent of the strain or stress level used in the experiment. Conversely,
once the modulus G(t) is known, the stress o(t) for
any strain y and time t is
= o(t)/y. Most polymers exhibit linearity
known through the relationship G(t)
at low stressesor strains (e.g., y z 0.5%) (36). In the case of food materials, strain
levels less than 3% are considered acceptable
to assure linearity(12). However, it
is important to examine the linearity over the entire range
of experimental condiso that the entire data set can be analyzed
tions and select the smallest strain level
uniformly.
D. ViscoelasticBehavior in DynamicTests
Although it is possible to conduct small-amplitude oscillatory tests i n tension
and compression, and there are commercial rheometers for such configurations
(37,38),thesheardeformationhasbeen
thedominantmode
of deformation.
Hence, the abbreviation SAOS, for small-amplitude oscillatory shear, is commonly used to represent dynamic viscoelastic tests. Concurrent with the developSAOS methodhasgainedmuchpopularity
ment of modernrheometers,the
among food researchers.
The SAOS test is based on applying a sinusoidal strain and measuring the
resulting stress. Of course, one can apply sinusoidal stress and measure the resulting strain. Either method,in principle, should produce the same material propertiesprovidedthattheimposeddeformation
is withinthelinearviscoelastic
region.
Let us consider Fig. 4, where a thin disk of a Hookean solid is subjected
to a sinusoidal shear strain r(t) between parallel plates
of the instrument:
r(t) = yo sin(ot)
(14)
where yo is the shear strain amplitude, o is the angular frequency, and t is the
time. Inserting Eq. (14) into Eq. ( I ) , the constitutive equation for a Hookean
solid, for y will give the resultant stress o (t), which is also sinusoidal:
o(t) = oosin(mt)
(15)
297
Hookean Solid
T
...................
....__ .__.
.......
."....... ........ ................

...... . q ; o f""
/=+
,
ot
1 Stress output
Fig. 4 Illustration of parallel plate geometry in a dynamic rheometer (top) and typical
stress-strain data for a Hookean solid (bottom). Stress and strain are in phase.
where o0is the shear stress amplitude. Therefore, for Hookean solids the resultant
stress wave is exactly in phasewith the input strain wave. This means that when
the strain is at maximum, so is the resultant stress, as shown in Fig. 4.
Let us now apply the same strain input [Eq.(14)], this time using a concentric cylinder configuration (Fig. 5 ) , to a Newtonian liquid whose constitutive equation is given, in simple form, by Eq. (2). Taking the time derivative
of the input strain function [Eq. (14)] and inserting the result for p in Q. (2)
yields:
o(t) = qmy, cos(0t)
(16)
298
Ak and Gunasekaran
Newtonian Liquid
Fig. 5 Illustration of concentric cylinder geometry in a dynamic rheometer (top) and

typical stress-strain data for a Newtonian liquid (bottom). Stress lags the applied strain
by an angle 6 (=7c/2).
As shown in Figure 5 , the stress response of a Newtonian liquid is exactly 90"

(or x12 radians) out of phase with the strain input.
A viscoelasticmaterialwithbothsolid-likeandliquid-likeproperties
should then exhibit an intermediate behavior (Fig. 6). The stress response of a
linear viscoelastic material to a sinusoidal strain input is given as (39):
o(t) = yoG'(o) sin(ot)
+ yoG"(o) cos(ot)
(17)
where two frequency-dependent functions,G' and G", represent the shear elastic
(storage) modulus and the shear viscous (loss) modulus, respectively. The term
in Eq. 17 having sin (cot) is in phase with the strain input, and the term having
cos(ot) is 90" out of phase with the strain input. Relative magnitudes of these
299
Viscoelastic Material
'
Stress output
Fig. 6 Typical stress-strain data obtained

in a dynamic rheometerfor a viscoelastic material. Stress lags the applied strain byan angle 6, (0 < 6 < d 2 ) .
terms will determine the degree of viscoelastic behavior. Therefore, we can rewrite the stress response of a linear viscoelastic material in a format similar to
the strain input as:
o(t) =
(so
sin(wt
+ 6)
(18)
where 6 is the phase angle between stress and strain.

The phase angle 6 varies
between 0 and 90" (or between 0 and 7cI2 radians), depending upon the relative
magnitudes of the two terms in Eq. (17).
Expanding the stress equation [Eq.(18)] using trigonometric relations (40),
we get:
o(t) = ((so c o d ) sin(wt)
+ (o0cos(wt)
sin6)
(19)
By comparing Eqs. 17 and 19 we obtain:
and
G' is a measure of the energy stored and subsequently released per cycle
of deformation per unit volume. It is the property that relates to the molecular
events of elastic nature. G is a measure of the energy dissipated as heat per
cycle of deformation per unit volume. G" is the property that relatesto the molecular events of viscous nature.
300
Ak and Gunasekaran
From the last two equations, we obtain another commonly used dynamic
viscoelastic property, the loss tangent (tan 6):
Hence, tan 6 denotes relative effectsof viscous and elastic componentsin a viscoelastic behavior.
Recall that for a Hookean solid 6 = 0; hence in-phase shear storage modulus is G = o&, and out-of-phase shear loss modulus is G = 0, as expected
6 = d 2 ; hence G =
and required. On the other hand, for a Newtonian liquid
0 and G = o,,/y,,, as expected and required.
Another way to present results of dynamic measurements is to use complex
modulus G*, the magnitudeof which is related to G and G through the following
equations (39):
Structurally, these terms can be explained

in terms of the ability of the
components to store (G) or lose (G) energy per test cycle. Thus tan 6 can be
viewed as the ratio of energy lost to energy stored per cycle and would be expected to decrease in samples with more and stronger elastic components but
would be higher for more viscous samples.
Similarly, complex viscosity (q*) can also be defined in terms of real (q)
and imaginary (q) parts of viscosity.
where,
The quantities G, G and q, q enable the rheological characterization of

a viscoelastic Inaterial on the basis of SAOS measurements. For instance, for a
Maxwell liquid the appropriate formulas may be derived using the above SAOS
relations as (18):
tic
Linear
301
Thus, tan 6 = I/or,enables calculating the relaxation time

'tR from the
SAOS data of G' and G". Furthermore, as o + 0, the dynamic viscosity q' +
q. Thus, theviscosity q occurring in the Maxwell model may be determined

from the dynamic results (18).
Often different publications denote the various viscoelastic parameters by
slightly different names and with different symbols and units.
It is desirable to
follow a standard set of nomenclature such as the one adopted by the Society of
Rheology (Table I).
1. Linear Range in Dynamic Tests

The linear viscoelastic region, where stress and strain are proportional and the
theory described in the previous section is applicable, can be quite different depending on the material and experimental conditions.
Table 1 Society of RheologyNomenclature for LinearViscoelasticity in Simple
Shear
parameter Rheological
Shear strain
Shear modulus (modulus of rigidity)
Shear relaxation modulus
Shear compliance
Shear creep compliance
Equilibrium shear compliance
Steady-state shcar compliancc
Complex viscosity
Dynamic viscosity
Out-of-phase component of q*
Complex shear modulus
Shear storage modulus
Shear loss modulus
Complex shear compliance
Shear storage compliance
Shear loss compliance
SI units
-
Pa
Pa
Pa '
Pa- '
Pa- I
Pa- I
Pa . s
Pa . s
Pa s
Pa
Pa
Pa
Pa- '
Pa I
Pa- '
Ak and Gunasekaran
302
Generally speaking, an investigation employing the SAOS method begins

with the determination of strainor stress limit for which the linear viscoelasticity
theory is applicable. For this, a strain or stress sweep test must be performed
unless, of course, this informationis already available in the literature. However,
before using the published values of linear limits, it is important to make sure
in the published
that the test material and test conditions match closely those
report.
A strain sweep test is conducted at a constant frequency (e.g.,
1 Hz) by
increasing the amplitude of the imposed strain. This test is schematically illustrated in Fig. 7. Once the linear viscoelasticlimit is determined, one can proceed
oKj
'r\A
0 05
0.01
0
O
'
K
; J
-0.01
-0.05
-0.02
3.14
Time (s)
6.28
3.14
Time (s)
6.28
3.14
Time (s)
6.28
STRAIN AMPLITUDE INCREASES
A
G',G
"*
4 -
,
0
k4
LINEAR
VISCOELASTIC
REGION
.*
>'
b
NONLINEAR
VISCOELASTIC
REGION
STRAIN AMPLITUDE
Fig. 7 Typical strain sweep test is conducted by applying increasing strain amplitudes
(top) and the resultant moduli (bottom) values are examined
for their departurefrom being
horizontal.
tic
Linear
Methods
303
withfurtherexperiments
(e.g., frequency-sweep)performedatstrains
(or
stresses) smaller thanthe limit. It is highly recommended to repeat the strain
sweep test at extremes of experimental variables, for there are data indicating
that the linear viscoelastic region can vary with test frequency, temperature, and
sample age (42.43). Variation of the linear viscoelastic region with test temperature is shown for mozzarella cheese in Fig. 8.
2. Features of SAOS Method

Recent investigations on food rheology often use dynamic mechanical measurements (e.g., SAOS) to study viscoelastic behavior of foods. The small amplitude
dynamic measurements are important and useful for many reasons.
I.
It is a nondestructive technique enabling measurements

to be made
while preserving specimen structure. This allows researchersto relate
dynamic rheological parameters to the molecular structure of a mate-
Shear Strain
Fig. 8 Linear viscoelastic region for mozzarella cheese as a function of cheese temperature. (Data from Ref. 43.)
304
Ak and Gunasekaran
2.
3.
4.
5.
6.
7.
8.
rial. These tests provide very

a
sensitive meansof studying the molecular motions that give rise to the phenomenon like glass transition
(44). This is discussed in detail later in Sec. 111.
It allows selective probing of molecular events by choosing a proper
frequency range since,in a dynamic mechanical experiment,the stress
response is dominated by relaxation processes with time constants
near I/o(45).
The linear viscoelastic region can be determined easily
in dynamic
testing by changing the amplitudeof the input strainor stress function
(46).
Two quantities,usually G and G, aremeasuredsimultaneously,
which gives a wayto check on both experimental error and applicability of time-temperature superposition (46).
When applicable, data from oscillatory tests can be utilized in timetemperature superposition technique to expand the frequency range,
whichotherwisewouldbeinaccessibleexperimentally.
The timetemperature superposition technique is based on the assumption that
a certain change of temperature changes the rate of all relaxation and
retardation processes by the same factor (47). The resulting graph of
the superposition procedure is called the master curve. Subramanian
and Gunasekaran (2 1) presented master curves for mozzarella cheeses
of differing composition and age. The superposition technique can
also be used with transient data
from linear viscoelastic creep and
relaxation tests. However, dynamic measurements are easier to perform with current rheometers than with transientstep strain and stress
experiments. Winter (45) stated that it might actually be more accurate to determine relaxation modulus G(t) from G and G data than
to measure it directly in a step-strain experiment.
Knowledge of dynamic properties such asG and G allows computation of all other linear viscoelastic properties as well as the material
behavior in other types of deformations such as tension (29,39). Much
effort has been, and continues to be, expended
in developing computer algorithms to perform such transformations accurately and fast.
Modern commercial rheometers often come with softwares that perform some of these transformation functions.
Molecular weight distribution of a material critically affects its propin its performance. Dynamic
erties and thus plays an important role
rheological data can be usedto determine the molecular weight distribution of materials (48-50).
It is generally faster to perform oscillatory tests than to perform other
line:lr Viscoelasticexperimentssuchas
creeD andrelaxation.This,
astic
Linear
Methods
305
however, depends on the number and range of frequencies required

(46).
9. Suddenchange ofthedisplacement(stressrelaxationtest)
or load
(creep test) is not required in oscillatory shear experiments (51).
10. Since it is afrequencydomainratherthanatimedomaintest,the
amplitude of the deformation yo and the time scale l / w can be independently varied (5 I).
These features of SAOS method are widely
used to investigate various
phenomena in different kinds of food systems as illustratedin the following section.
111.
APPLICATIONS
A. Transient Tests
In terms of experimentalmethodology,compressiontesting
is anexpedient
it provides
technique to distinguishgoodandbadsamples.Butbeingstatic,
limited information on the mechanism(s)of the process(es) involved, understanding of which is needed for the development of newer products and/or textures.
Transienttestsareinherentlyusefulforevaluatingmaterialstructureand/or
structure development during processing. This has been the primary application
of transient tests. Nonetheless, there have been many investigations
to develop
empirical correlations between the transient test parameters (relaxation time. retardation time, equilibrium modulus, etc.) and some food quality/texture attributes.
Transient tests have been very popular among food researchers because of
their simple test configuration. Stress relaxation tests can be readily performed
using universal testing machines
(e.g., Instron, Texture Analyzer). Creep
tests
can be performed by means of a loading device with a provision to continuously
monitor (and record) the sample deformation. However, some practical problems
exist. In the case of stress relaxation tests, it is virtually impossible to impose
an instantaneous strain, andin the case of creep tests, dueto continually changing
sample cross-sectional area, the applied stressis often not a constant, as generally
required by the models used for data analysis.
Evageliou et al. (52) described developing a new fat mimetic by carefully
a
evaluating the structure of the material by creep tests. A model combining
Maxwell element in series with one or more Voigt elements is used to describe
a s protein
time dependency of the linear creep compliance. Biopolymers such
(from egg, milk. gelatin), intact and modified starch, and/or gums
(e.g.. carrageenan, alginate, pectin) are used to provide structure to fat mimetics. Since no
306
Ak and Gunasekaran
biopolymer by itself can provide the required structure and a plastic flow,
mixtures are used(e.g., gelatin and maltodextrin)in an attempt to impart plasticity. However, if incompatible polymers are mixed, phase separation leads to
an unacceptable product (52).
Ingel systems,networksareformedbyphysicalcross-linkswith
finite
lifetimes, which will stretch and eventually relax when under mechanical stress,
thus allowing the system to flow. Creep curves of full-fat products suchas butter
and margarine exhibit lower initial strain (about two orders of magnitude) than
those of biopolymer networks (37,52,53).The difference is related to permanency
of cross-links. The fat lattice is formed by the short-range van der Waals forces
of attraction between glyceride crystals as opposed to the long junction zonesof
helical strands in a biopolymer network.
Comparing the creep curves of margarine and gelatin gels, the fat crystals
were reported to flow more rapidly (larger compliance) than did the mimetic
made of gelatin triple helix. Much of the creep was observed to occur beyond
the instantaneous deformation. The instantaneous compliance is associated with
of a network, and the higher proportion
of the
the initial elastic deformation
total rearrangement in gelatin gels reflects the elastic character of gelatin helical
associations as opposed to the plasticity of a liquid/solid fat body (54).
Sakurai and Nevins (55) recommended use of stress relaxation measurements to study fruit and vegetables since physiconlechanical parameters derived
through application of the Maxwell model were related to time-dependent biochemical and microstructural changes associated with softening (56). The minimum stress relaxation time for tomatoes changed with ripening as a function of
(56) suggested that
softening and position within the fruit. Sakurai and Nevins
stress relaxation tests might be able to distinguish differences in fruit tissue due
to ripening that are not observable by means of large deformation tests.
Creep tests modeled as per the Burgers model (series combination of the
MaxwellandKelvinmodels)mayprovidemoreinformationthanthesimple
stress-relaxation model (Maxwell). Creep tests allow prediction of elastic, viscoelastic, and viscous flow characteristics separately. The creep test is perhaps the
most appropriate means to determine the elastic parameter (instantaneous compliance) of a material (57). Model parameters may be associated with discrete components or constituents of the material being tested (14,15,58). This facilitates
of the physicomechanigreater understanding of the microstructural determinants
cal behavior or texture exhibited
(1 2). A six-element model is found to best define
in terms of physicomechanical
the viscoelastic behavior of tomato pericarp tissue
interpretation of the tissue softening mechanisms (12). The instantaneous elastic
properties of tissue are attributed to the combination of cell turgor pressure and
primary cell wall strength as dictatedby cellulose. The retarded viscoelastic properties are related to the independent changesin hemicelluloses and polyuronides.
The steady-state viscous flow properties are related to higher cell wall fluidity
307
arising from a molecular weight downshift

in cell wall and/or middle lamellar
polymers and from exosmosis. The loss of turgor, breakdown of polyuronides,
and overall increase in cell wall fluidity were each found to contribute about 2530% to the apparent softening of tissue. The downshift in molecular weight size
distribution of hemicelluloses was presumed to contribute 10-15% to the tissue
softening during ripening. Some recent transient viscoelastic investigations
on
raisins indicate that the raisin skin is the primary elastic element exhibiting relaxation times longer than those for theflesh, which contributes to the viscous character (59,60).
Campanella and Peleg (6 1 ) investigated stress relaxation properties of processed American cheese in conjunction with mechanical models (three-element
solid model) to simulate the interaction of the cheese with tongue or fingers.
Though they made many oversimplified assumptions, they demonstrated that the
sensory system performance could be analyzedin mechanical terms. Purkayastha
et al. (62) fitted the creep curves of Cheddar cheese and potato flesh by a fourof the two materialsis clearly
parameter model. The general rheological behavior
expressed in terms of the model constants. Ma et al. (63) studied the viscoelastic
properties of cheeses using creep and dynamic oscillatory tests. The creep test
data are used to identify the differences in viscoelastic properties of cheeses due
to fat reduction and addition of lecithin. It is shown by creep recovery measureof granular
ments on the reduced-fat processed cheese that adding a small amount
of the
soylecithin or hydrogenatedsoylecithinimprovestexturalproperties
cheese without affecting sensory acceptance scores (64).
Kuo et al. (65) used a six-element Kelvin
model to describe a modified
squeeze flowtest data to determine cheese meltability. Accordingly, the following
equation is used to represent the model for the cheese flow data:
where J(t) is the total creep compliance at time t, J,, is the instantaneous rigidity
compliance, J , and J2 are the retarded compliances, T~ and zz are the retardation
times, and vv is the Newtonian viscosity.
Thevaluesoftheviscoelasticparameterscalculatedfor
bothfull-and
reduced-fat cheeses are given in Table 2 . The reduced-fat cheese had a lower
instantaneous compliance (J,,) than the full-fat cheese. This suggeststhat a reduction in fat resulted in an increase in the elastic (or solid-like) character of the
cheese. The instantaneous compliance may be related to the undisturbed protein
(q,)of reduced-fat cheese
network structure (63). The higher Newtonian viscosity
suggested a greater resistance to flow at longer time. Newtonian viscosity may
be attributed to the breakdown of protein network (66). Thus, the reduced-fat
cheese could be considered to retain more of its solid-like viscoelastic structure
than the full-fat cheese.
ters
308
Ak and Gunasekaran
Table 2 Six-ElementKelvinModelParameters of Regular-FatandReduced-Fat

Cheddar Cheeses Subjected to a Creep Test for Meltability Evaluation
Cheddar cheese"
Six-element Kelvin
model
Jo( 1 /kPa)b
J , (1 /kPa)'
Jz( 1/kPa)'
0. I 4x
0.27"
0.55'
3.35"
28.04'
137.9'
2, (S)d
T?(s)d
qv(kPa.s)'
*Differentletters (x,y) In eachrowindicatesignificantdifference

values.
Instantaneous compliance.
Retarded compliance.
Retardation time.
Newton~an viscosity.
Source: DatafromRef. 65.
0.1 1"
0. 19x
0.46x
3.25"
23.89"
149.6"
( p
< 0.05) betweenthemean
A typical creep curve corresponding to the six-element model used by Kuo

et ai. (65) is shown in Fig. 9. The curve has three segments corresponding to the
Hookean, KelvinNoigt, and viscous elements. The retarded compliances (J, and
J2) represent the principal components of the viscoelastic behavior of Cheddar
cheeses. This reflected a high degree of retarded KelvinNoigt-type deformation
.........
0.9
"=
:,-:
........ ..~"."."."._..
Wl")
J,+J,
...........................................
" ' . ,
50
100
Jo
150
Time (s)
Fig. 9 Six-elementKelvinmodelusedtocharacterizecheesemeltability.(Datafrom
Ref. 65.)
astic
Linear
309
in Cheddar cheese under external loading. The instantaneous slope of the creep
curve is calculated by taking the first derivative of Eq. (31) at time zero. This
instantaneous slope is defined as the viscoelasticity index (VEI), which is computed as follows:
Cheese meltability determined by anothertest (67) and VEI correlated well,

indicating the applicability of the creep test as a practical means for describing
cheese meltability.
Stickiness is a technologically important textural property that may cause
major processing problems in baking and confectionery industries. Recently,Hoseney and Smewing (68) defined stickiness as the force of adhesion that results
when two surfaces are contacted with each other. In the same article, they also
discussed problems associated with measuring stickiness and described various
instrumental methods used to measure food stickiness. Stickiness
of wheat dough
has been determined by the standard peel test, and the transient and dynamic
rheological tests (69). The stress relaxation gradients were observed to be greater
for sticky doughs than nonsticky doughs. Moreover, the similarity of stress relaxation gradients to the peel-rate plots is taken to indicate that adhesion is mainly
a function of the bulk viscoelastic properties of the dough (69).
B. Dynamic Tests
1. Gelation and Properties of Dairy Products
One commonuse of the SAOS technique is to measure linear viscoelastic properties of materials during gelation and aging (or curing). Here, we consider examples of milk gels formed by rennet or acid addition, and other dairy products (e.g.,
yogurt, cheese). Studies on gel development usually employ concentric cylinder
system, whereas those on mature gels and cheese use parallel plate or cone and
plate configuration. The cone and plate system
is similar to theparallelplate
configuration (Fig. 4) except that the top plate is replaced with a cone of small
angle.
The SAOS method has been used to follow gelation kinetics of various
proteins and polysaccharides. In these studies, the technique is used to monitor
the change in the dynamic moduli G and G with time, and in some cases with
temperature, at some selected frequency (usually 1 Hz). In a gelling system we
usually observe sudden increases in G and G after an initial lag period. This
is often followed by a gradual decrease in rate of increase of G and G, which
eventually levels off in most cases. Moreover, once the mature (or fully cured)
gel is obtained, it is common practice to make further measurements such as (a)
Ak and Gunasekaran
310
strain sweep (G and G vs. y) to determine the linear viscoelastic region, (b)
frequency sweep (G and G vs. O ) to determine the elastic character of the gel,
and (c) temperature sweep (G and G vs. T) to evaluate thermal characteristics.
Oneimmediateconcernariseswhen
we want to determineviscoelastic
to elastic gel. This is whether
properties of a system going from viscous milk
the applied strain or stress amplitude affects the aggregation process. Dejmek
(70) measured G and G during rennet coagulation of milk. When the applied
shear strain was high (i.e.,yo = 0.2), the resultingG* from the continuous oscillationmodewassignificantlygreaterthan
that from the intermittent oscillation
mode. Apparently, continuous oscillation affects the aggregation process. Dejmek
also reported that the coagulation process was not affected by continuous shear
strains below 0.05 at 1 Hz (70). To avoid disturbing the skim milk gel network,
Zoon et al. (71) started oscillatory shear measurements
at 1 rad/s (-0.16 Hz)
only after a weak gel formed (i.e. G = 2 Pa).
Dejmek (70) reported that the linear viscoelastic strain limit for mature
rennet gels is 0.05. The corresponding limit for rennet-induced skim milk gels
is reported to be 0.03 (70). Paulsson et al. (72) applied strain amplitude of 0.02
of
at 1 Hz to stay in the linear viscoelastic region while monitoring formation
heat-induced P-lactoglobulin gels. Most studies on dairy gels seem to use strain
amplitudes less than 0.05 and a frequency of I Hz.
A typical curve obtained during gelation (or curing) is schematically illustrated in Fig. 10. The beginning of time axis (t = 0) usually refers to the addition
Time
Fig. 10 Illustration of typical dynamic moduli data (G and G) as a function of time

during gelation of milk showing the initial lag period and the subsequent gel development
period.
311
of rennet or acid to the milk. There is generally a lag period before dynamic
properties attain values greater than the minimum torque of the rheometer. After
the initial lag period, both G and G increase with a rate that depends on the
experimental conditions. Whether a dynamic property reach a plateau value
or
not also depends upon experimental conditions such as the temperature at which
the gel is formed and aged (73).
Bohlin et al. (74) designed and developed a dynamic testing instrument
with coaxial cylinder geometry and used it to monitor coagulation of milk under
conditions similar to thosein cheesemaking. They investigated effectsof calcium
on coagulation of
chlorideandrennetconcentrationsalongwithtemperature
milk. Their data show that higher concentrations of rennet and calcium chloride
and elevated temperatures result in an earlier start of coagulation and a faster
gelation rate, leading to stiffer gels (i.e., higher G* values). The start of coagulation is identitied as the time at which G and G began to deviate from zero (see
Fig. IO). Nakamura and Niki (75) studied the influence of calcium concentration
on rheological properties of casein micelles during gelation. They concludedthat
of
changing calcium concentration affects gelation rate but not the mechanism
gelation.
Several quantitative coagulation parameters may be obtained from gel development curves. For instance, Dejmek (70) presenteda method based on threeto derive coagulation parameters such as
parameter Scott Blair-Burnett model
of incipient gelation from
the time constant of gel build-up and time constant
gel development curves. Zoon et al. (73) obtained clotting time as the time from
rennet addition to the formation of visible clots; it generally decreased with increasing gelation temperature. On the other hand, van Hekken and Strange (76)
determined coagulation time as the time from rennet addition to the point where
et al. (77) conducted
G exceeded G, or tan 6 became less than one. Guinee
SAOS tests to investigate the effect of heat treatments of milk before ultrafiltration on the coagulation characteristics of retentates with different protein levels.
They also used the Scott Blair-Burnett model to extract several coagulation parameters, which enabled them to compare the effects quantitatively.
SAOS measurements are also used to characterize behavior of dairy products (e.g., yogurt, cheese, whey gels) as well toasevaluate effectsof many factors
on viscoelastic properties of these products. Skriver (78) conducted SAOS tests
on stirred yogurt to assess the effect of technological parameters such as culture
type, fermentation temperature, and dry matter content on rheological characteristics of the product. Based on dynamic shear data, stirred yogurt is characterized
as a weak gel. The main features of weak gels are given by Ross-Murphy (79)
as G > G. both G and G are largely independent of frequency, and the linear
viscoelastic strain limit is small (y < 0.05). Moreover, Skriver (78) found that
exo-polysaccharide produced by the ropy culture did not contribute
to the dyin contrast to the viscometry
namic gel stiffnessof stirred yogurt. This tinding was
312
Gunasekaran
Ak and
(large deformations) results where the contribution

of exo-polysaccharide to shear
stress is reported to be significant (80).
Ronnegird and Dejmek (81) studied the development of gel structure in
G data for the
set yogurt by oscillatory shear measurements and compared the
to find that modulus of the
set yoghurt with that for commercial stirred yogurt
latter is about 10 times smaller.
Paulsson et al. (72) studied gelation of heat-induced P-lactoglobulin
by
dynamic rheometry at different pH levels (4.5, 5 , 7) and protein concentrations
(3, 4, 5% mass/vol). They reported that the temperature at the start of gelation
of G* is
is mostly independent of pH and protein concentration but the value
influenced mainly by the protein concentration and to a lesser degree by the pH.
It is interesting to note that the exponent n
in the relation JG*Jr. (cy, where c
is the protein concentration, varied between 2.2 and 2.6 for rennet and acid milk
gels (7 1 ) as well as heat-induced P-lactoglobulin gels (72) when pH was below
7. A higher value for n is reported when the pH was 7 (72).
Viscoelastic properties of acid casein gels made by slow acidification with
glucono-&lactone (GDL) have also been determined using oscillatory shear tests
(82,83).Somesimilaritiesanddifferencesarenotedbetweenrennet-induced
(skim) milk gels (71,73) and acid casein gels made with
GDL (83). Forboth
types of gels, gelation time increased with decreasing gelation temperature. The
gelation time of GDL-induced gels is about 5-15 times that of rennet gels depending on the gelation temperature. The G of gels with GDL reached plateau
values of 500-600,100-200and
<20 Pa at 20, 30, and 4 0 C respectively,
whereas the G of rennet gels was still increasing after 72 hours of aging at 20C
and reached plateau values of 125 and 75 Pa at 30 and 4 0 C respectively.
Salvador and Fiszman (84) used dynamic rheometry to study properties of
acidic and neutral milk gels produced by gelation
of commercial gelatin. I t is
by gelatin than
found that the values of G were higher for milk gels formed
those for aqueous gelatin gels. Thisresult is taken to indicate stabilization of the
gel network by milk components (84).
Navarro et al. (85) recently reported on the linear viscoelastic properties
of commercial samples of D u k e de Leche (a typical Argentine sweet spread)
with different compositions. The linear viscoelastic limit for these sweet spreads
varied between 3.5 and 9%, the larger value belonging to the confectionery type,
which contained agar as thickener. Dynamic rheological data indicated that the
behavior of the common and low-calorie Duke de Leche was more like that of
the concentrated solutions and dispersions, whereas the behavior of the confectionery type was similar to that of a weak gel (85).
Another use of the SAOS test is demonstrated by Ozer et al. (86) where
rheologicalproperties of six differentlabneh(concentratedyogurt)samples
were compared using data from SAOS, penetrometer, and viscosity tests. The
destructive penetrometer and viscosity measurements failed
to reveal expected
Viscoelastic Linear
313
differencesamonglabnehsamplespreparedafterdifferenttreatments.However, oscillatory shear tests clearly differentiated between the samples and, therefore, are considered to be more reliable to determine rheological properties of
labneh.
the qualityof prepared foods
Melting characteristicsof cheese are critical to
containing cheese (e.g., pizza). Several methods have been developedto measure
meltability of cheese (87). There is, however, a lack of correlation among traditional methods (88), and thus new methods are needed. Riiegg et al. (87) stated
6, could be potentially used as a predictor for cheese
that the loss tangent, tan
meltability. Indeed, the use of dynamic rheometry to study cheese properties at
high temperatures has been increasing (89-91).
(G', G",
Ustunol et al. (89) correlated the dynamic rheological properties
and G*) of Cheddar cheese up to 90C with the meltability data from the traditional Arnott test. The minimum complex modulus G* is suggested as a possible
meltability indicator as it showed a significant correlation ( I = -0.80) with the
empirical meltability results. A recent study on the melt characteristics of imitation cheese reports a higher correlation coefficient ( r = 0.96) between the maximum tan 6 values and meltability results obtained by the empirical Olson-Price
tube method (90).
Meltability of process Cheddar cheese made with different emulsifying
salts has been measured by dynamic stress rheometry (91). In this study, the
parameter to compare behavior of cheeses was the so-called transition temperature, which is defined as the lowest temperature at which tan 6 = 1. The change
from more elastic to more viscous behavior occurs at a lower temperature for
process cheese containing trisodium citrate (56.5"C) than that containing disodium phosphate (64.6"C) (91).
A significant concern in dynamic rheometry at high temperatures
is the
slippage (92), which results in inconsistency in the rheological data (93). Howto the plates and/or using serrated
ever, remedies such as bonding the sample
plates or compressing the specimen slightly have
been applied to reduce the slippage problem (89-91,94-96). Another technique
to study properties of melted
cheese, which does not suffer from slippage problem, is the lubricated squeezing
flow. This technique is a simple but fundamental alternative to the traditional
empirical meltability tests and has been frequently usedto study cheese meltability (67,97-99). The sample and test geometry of the squeeze flow method makes
it also suitable for performing creep and stress relaxation tests within linear viscoelastic region (65).
2. Determining Gel Point in Polymeric Systems
Another area of research where SAOS method has found wide

use is in determining gel point, i.e., the moment at which a polymer/biopolymer system changes
314
Ak and Gunasekaran
from a viscous liquid(sol) to an elastic solid (gel) during the course

of the gelation
process.
Determining the gel point from rheological properties such as steady shear
viscosity fortheliquidstateandequilibriumshearmodulusforthesolid
state requires extrapolation (Fig. 1 1 ) and suffers from singularity at the transition (100). SAOS method,however,providescontinuousrheologicaldata
for the entire gelation or curing process. Hence, SAOS has become widely used
of maturegels (100to investigatethesol/geltransitionandtheproperties
104).
Ross-Murphy (105) listed a number of rheological means to detect gel point
or gel time: (a) when the signal from gelling system becomes
just greater than
the background noise, (b) whenG becomes higher than a chosen threshold value,
(c) when G becomes just greater than G (the cross-over method), and (d) when
tan 6 becomes independent of frequency (the Winter-Chambon method). Method
I suffers from being dependent on the instruments minimum torque value, which
may vary among commercial rheometers. Method 2 requires prior knowledge of
3 , as Tung and Dynes(106) pointed out,
the proper value for gel strength. Method
depends on the frequency of the oscillation test. Method4, based on fundamental
arguments that have been experimentally supported, appears to be the preferred
way in recent studies to detect gel point.
Fig. 11 Illustration of discontinuityinthesteadyshearviscosity/equilibriummodulus

data at gel point also known as the critical point (t,) during sol-gel transition (Adapted
fromRef.100.)
315
Viscoelastic Linear
In the study by Tung and Dynes (106), it was suggested that the time at
which G and G curves cross each other can be used for determining the gel
point. This studyalso reported that the gel time determined as such was a function
of frequency of the oscillatory test. Recognizing that the gelation of a material
must not depend on the frequency of the rheological test, Winter and Chambon
(100) havedeveloped a newmethodforlocatinggelpoint.According
to the
Winter-Chambon criterion, tan 6 (= G/G) at the gel point becomes independent
of frequency (104) (Fig. 12). Moreover, the cross-over method is a special case
of the Winter-Chambon method (102,104).
Lopes da Silva and Gonqalves (107) studied rheological properties of curing high-methoxyl pectin/sucrose gels at different temperatures using SAOS experiments. The G-G cross-over method could not be used asa criterion to identify the gel point. They instead applied the Winter-Chambon criterion. Michon
et al.( 1 08) applied the Winter-Chambon criterion to determine critical parameters
of gelation (e.g., time, temperature) at different polymer concentrations for systems involving iota-carrageenan and gelatin. When gelatin was cooled from
60C
the gelling time was found to be 44 minutes. In another set of experiments, the
gelation temperature of iota-carrageenan was determined to be 53.5OC based on
the same criterion. These researchers provided phase diagrams where critical temperature of sol/gel transition was charted against concentration of the polymers.
These graphs showed that at a given temperature, a gelatin with mean molecular
mass of 70.000 daltons would require a much higher concentration for the sol/
gel transition than that with a mean molecular mass of 182,000 daltons.
.
4
c
(TI
.
I
-
0.1
Gelation time (min)

Fig. 12 At thegel point, loss tangent (tan 8 ) is independent of frequency (w). (Data
fromRef. 103.)
Gunasekaran
316
Ak and
Labropoulos and Hsu (109) have used the SAOS method to investigate gelforming ability of whey protein isolate (WPI) dispersions subjected to different
heattreatmentsandotherprocessingvariables(e.g.,pH,concentration).Employing the Winter-Chambon criterion for gel point detection, a wide range of
gelation times from 12 to 164 minutes is obtained depending on the experimental
conditions. Information from such studies is expected to allow processors to obtain desired gel properties by controlling the variables during gelation of WPI
dispersions.
3. Characterizing Food EmulsionGels

Different types of emulsions are widely encountered in several food products
and maysuch as ice cream, margarine, butter, beverages, sauces, salad dressings,
onnaise (1 10).The stability of food emulsions is arather important and complex
phenomenon that greatly influences the quality and
performance of a product.
Good emulsion stability means that the size distribution and the spatial arrangea
ment of droplets do not change significantly during the observation time. In
recent review article, Dalgleish
(1 11) discussed different types of instabilities
that may occur in food emulsions. The instabilities observedin oil-in-water emulsions are illustrated in Fig. 13.
Emulsion droDlets
may
Which
mav:
Cream
lead to:
A
Creaming
01
Flocculate
(seml-reversibly)
0
01
Aggregate
(irreversibly)
andlor
or
Coalescence
flocculation
or
Q0
Flocculate
by bridging
~~~~~
.
"
+
@-0
Fig. 13 Instabilities inanoil-and-water
emulsion. (Adaptedfrom Ref. 111.)
317
Rheological properties of food emulsions play important roles in the stability as well as texture and mouthfeel of these products. In this section we present
examples on the use of SAOS method to characterize food emulsion gels, to
understand stability mechanisms, and to determine the roles of various factors
contributing to the emulsion stability.
Dickinson and Golding (1 12) used the SAOS method (0.04 Pa, 0.1 Hz) to
examine viscoelastic changes due to development of a particle network in emulsions prepared at pH 6.8 with sodium caseinate as the only emulsifier( I -6% by
mass) and n-tetradecane as the dispersed phase (10, 35, or 45% by volume). The
2% (by mass) caseinate emulsion (35% by volume n-tetradecane, pH 6.8, 30C)
behaved like a viscous liquid with G > G at all times. This behavior is associated with a nonflocculated system.
In contrast, for the 6% by mass caseinate
emulsion (35% volume n-tetradecane, pH 6.S,3O0C), there was an initial decrease
in the dynamic moduli followed
by a steady-state region, which was followed
by a steady increase of both quantities (G faster thanG) until a cross-over point
(G = G). Thisbehavior is associatedwiththe
floc reorganizationtomore
closely packed structures and the formation of a gel network by the flocculated
particles.
Muiioz and Sherman( 1 13) measured the viscoelastic properties of commerSAOS tests
cial mayonnaise, reduced-calorie mayonnaise, and salad creams using
with a controlled stress rheometer (1 Hz, 8.96-1 8.51 Pa). They have explained
the observed differences in viscoelastic properties by variations in the ingredients
of these emulsions. For instance, the lower
G value of one mayonnaise compared
to another is related to the presence of sugar in the former, as sugar molecules
exerted a shielding effect on protein groups involved in interaction and network
formation among the oil droplets. The lowest G values for salad creams are
attributed to the lower oil content, yielding a lower concentration of oil droplets
in the emulsion. This is in accord with the results of Dickinson and Golding
( 1 12) in that even when protein concentrationwas high (>8% by mass), emulsion
made with lower volume fraction of dispersed phase (oil) had a weaker structure.
Other studies by Gallegos et al. ( 1 14) and Ma and Barbosa-Cinovas ( I 15) confirm that higher oil contents result in higher dynamic moduli for commercial and
modelmayonnaisesamples.Moreover,thestrainsweepmeasurements
at I O
rad/s on model mayonnaise samples indicated a narrower linear viscoelastic region for samples with lower oil content ( 1 14).
Past work on whey proteinshas shown that co-homogenizationor preemulsification of the oil with lecithin reduces significantly the strength of heat-set
emulsion gels ( 1 16). Consequently, Dickinson and Yamamoto ( I 17) used the
SAOS method (1 Hz, maximum strain amplitude of0.5%) to examine thoroughly
the influence of lecithin added after emulsion formation on the properties of heatset P-lactoglobulin emulsion gels. The dynamic data(G and G) clearly showed
that the addition of lecithin at a concentration of 4.4% by mass increased G of
31 8
Ak and Gunasekaran
the heat-set @-lactoglobulin emulsion gel at least

10 times. It is suggested that
once P-lactoglobulin becomes adsorbed at the surface of oil droplets, the role of
lecithin is then toformcomplexes withadsorbedandfreeP-lactoglobulin
to
reinforce the heat-set emulsion gel network (1 17). The positive effect of lecithin
lesser extent.
is also observedin G of heat-set @-lactoglobulin gel (no oil) buta to
It is concluded that the lecithin-containing emulsion gels behaved more like a
strong gel since G is less frequency-dependent in the range 10 to 2 Hz as
compared to the @-lactoglobulin gels made without lecithin.
to determine effects of proDynamic measurements are also employed
cessing parameters, such as emulsification temperature and machine type, on the
stability of salad dressing emulsions (1 18). Franco et al. ( I 18) reported that the
linear viscoelastic region of emulsions is significantly affected by the processing
parameters. Moreover, their data indicated that higher energy input during emulsification and elevated processing temperature (50C) result in higher values of
dynamic moduli because of enhanced network formation of flocculated oil droplets. This in turn improved the stability of the emulsions.
4. DeterminingSensoryPerception
Linear viscoelastic tests have also been applied to study the interaction between
texture, flavor, and taste. It has been shown that complex viscosities q* from
dynamic tests rather than steady shear viscosities give better correlation with
sensory thickness for weak gel systems such as liquid and semisolid foods( 1 19121). Hill et al. (122) investigated the relationship between the perceived (sensory) thickness, taste and flavor and rheological parameters
of a lemon pie tilling.
They established a general linear relationship between perceived thickness and
q* over theviscosityrange of 1,000-70,000mPa.s. (Fig. 14). Richardsonet
2.5 3.5 3.0
4.0
log (q, rnPa.s)
4.5
5.0
Fig. 14 Linear relationship between perceived sensory thickness and complex viscosity.
(Data fromRef. 122.)
Viscoelastic Linear
319
al. (1 19) covered the viscosity range of 10-10,000 mPa.s using model systems
thickened by xanthan, guar, and starch. These results indicate that q* measured
at a frequency of 50 rad/s can be considered a suitable criterion for predicting
perceived thickness in widely differing systems (122).
The relationship between taste and flavor and rheological properties seem
to depend on the structure of the solution and the concentration. It is well established that the intensity of perceived taste and flavor decreases as the product
thickness increases ( I 23). In model solutions containing random coil polysaccharides, suppression of taste and flavor starts when the polysaccharide concentration
exceeds the critical concentration where the polymer coil starts to entangle. The
exception to this is xanthan gum where good flavor release is achieved at a concentration in excess of the critical concentration ( 120,121,124,125). At comparable viscosities, cornstarch suppresses sweetness far less than random coil polysaccharide, which is thought to be a consequence of a weak gel structure ( 1 22).
Windhab et al. (126) established stability domains for emulsions processed
under various conditions using different rotor/stator gap geometries. They reported a linear relation between the sensory thickness impression from the spoon
test and the ratio of G/G from the dynamic rheometry for optimally processed
and stable emulsion samples.
R@nnet al. (127) recently reported on the predictionof sensory properties
from dynamic rheological measurements for low-fat spreads. Ten commercial
low-fat spreads are evaluated bothby a trained sensory panel and by SAOS measurements. Two of the I 3 properties evaluated by the panel, namely, meltability
of the temperature sweeps
and graininess, are successfully predicted from features
(G* vs. T) from the rheological measurements.
5. Determining GlassTransitionTemperature
Materials with amorphous or partially amorphous structures undergo a transition
from a glassy solid state to a rubbery viscous state at a material-specific temperature called the glass transition temperature,
Tg (29,38,44). Glass transition in
amorphous food materials generally occurs over a range
of temperature rather
than at a single temperature (44,128,129).
The technological importanceof the glass transition phenomenon for different types of foods has been discussed in detail in the literature (44,130-136).
The glass transition or the associated parameter Tg has a great effect
on the
processing,properties,quality,safety,andstabilityoffoods(137).
Tg or the
difference between the temperature of a material and
its Tg (Le., T-Tg) affects
the physical and textural properties of foods (e.g., stickiness, viscosity, brittleness, crispness, or crunchiness), the rates of deteriorative changes (e.g., enzymatic
reactions, nonenzymatic browning, oxidation), and the success of many processes
(e.g., flavor encapsulation, crystallization). Hence, the knowledge of Tg is essen-
320
Gunasekaran
Ak and
tial in assuring the quality, stability, and safety of various foods such as confectionery products, breakfast cereals, baked goods, coated flavors, frozen products,
and food powders (128,130,133,134,137-146).
At the glass transition temperature,* properties such as the thermal expansion coefficient, the dielectric constant (for polar materials), and the heat capacity
exhibit a discontinuity. Thus, techniques measuring such property changes have
of Tg (e.g., dilatometry, calorimebeen developed for experimental determination
try) (29,39). Differential scanning calorimetry (DSC) is probably the most commonly used technique for determining Tg. Besides DSC, the
utility and advantages of other techniques such as nuclear magnetic resonance (NMR), electron
spin resonance spectroscopy (ESR), and dynamic mechanical (thermal) analysis
(DMA or DMTA) have been increasingly appreciated (148-153). Kalichevsky
et al. (151) stated that the dynamic mechanical techniques are more sensitive
than DSC to the molecular motions around Tg and the relaxations below Tg.
When dynamic mechanical spectroscopyis employed within the linear viscoelastic regime to determine Tg, the storage and loss moduli (G' and G" or E'
and E"), and loss tangent (tan6 = G"/G' or = E"/E') are measured as a function
of temperature at a constant frequency and a selected heating or cooling rate. In
of biological and synthetic polymers exhibits
glass transition, the storage modulus
a sharp drop with temperature. whereas the
loss modulus or tan 6 shows a characteristic peak (29,39,152,154-157). The decrease observed in modulus of amorphous synthetic polymers is typically about 3 orders of magnitude (29), whereas
that observed in biopolymers is about one order of magnitude (148,151-153).
Moreover, as Peleg (158) demonstrated, the plot of stiffness or rigidity (E', G')
versus temperature in the transition region of biomaterials has a downward concavity, which cannot be described by conventional models, such as WLF ( 1 59),
developed for synthetic polymers. Peleg, therefore, suggested another model and
demonstrated its applicability to describe the stiffnessor rigidity versus temperature relationship of biopolymers at the transition region ( I 29,155,158,160- 162).
Different definitions have been used to facilitate experimental determination of Tg from the dynamic mechanical data: (a) temperature corresponding to
the onset of drop in storage modulus T,,,,,, (b) temperature corresponding to the
(c) temperamidpoint of the glass transition region for storage modulus Tlllldpolnl,
turewheretheextrapolatedline
of theinitialmodulusintersectsthat
of the
(d) temperature corresponding to the loss modulusG" peak
steepest slope TInlerrec,,
6 peak Tlans.These are schematiTG",and (e) temperature corresponding to the tan
of Tg for the same material
cally illustrated in Fig. 15. It must be noted that values
may differ slightly or significantly depending on the definition, the experimental
conditions (e.g., heating/coolingrate in DSC and test frequency in DMA or
* No phase change is involved as
the word "transition" may possibly ,imply (147).
321
Steepest slope
Initial slope
Moduli=0.5x Initial moduli
T
.,
'tan
Temperature
Fig. 15 Schematic drawing of different ways of obtaining glass transition temperature

from dynamic mechanical data.
322
Gunasekaran
Ak and
DMTA), and the technique (e.g., DSCvs. DMA) (29,153,154). Hagen et al. (163)
reported that Tg values, for instance, of an unfilled natural rubber differby about
10Cwhen determined by using two commercial dynamic mechanical instruments (DMA vs. DMTA).
Frequently the temperature corresponding to G or tan 6 peak is used as
a marker ofTg (29,148,149,153,164). However, Peleg (157) showed
by computer
6 peak location does
simulations and published experimental data that the tan
not always correspond to the transition zone even for the same material with
different moisture contents. Peleg (157) has instead suggested, as a more meaningful index of Tg, to use the temperature at which 50% of the initial stiffness
(i.e., storage modulus, G) is lost.
Several factors, such as composition,* molecular weight, and features
of
chemical structure (e.g., cross-links, side groups) can alter
the glass transition
temperature of materials. Among the constituents of foods, water is an effective
andubiquitousplasticizer,whichlowersthe
Tg of mostbiologicalmaterials
(134). Plasticizers are relatively low molecular weight materials
that, when added
it
to amorphous polymers, lead to a large increase
in mobility and thus make
easier for changes in molecular conformation to take place.
Cocero and Kokini (148) demonstrated the plasticizing (or softening) effect
of water, as measuredby the storage modulusG, on the major protein component
(i.e.. glutenin) in wheat Hour. Hallberg and Chinachoti (166) used DMAto study
phase transitions in shelf-stable MRE (meal, ready-to-eat) bread. Three distinct transitions are reportedin fresh MREbread and among those the main transition temperature decreased from about 160
to - 1 1C as the moisture content
increased from about 2.6 to 28.8%. It is also found that transition temperatures
( 166). Gontard et al.
( 167)
remained nearly constant throughout 3 years of storage
on mechanical
reported on the strong plasticizing effect of water and glycerol
and barrier properties of edible wheat gluten films. Kalichevsky and Blanshard
of amylopec(164) studied the effectof fructose and water on the glass transition
tin to find that the effect of fructose on the Tg of amylopectin is greater at lower
water contents.
The depression of Tg, due to plasticization of amorphous components by
water or other plasticizers, to the vicinity
of ambient temperatures may have a
significant effect on the shelflife and stability
of foods (128,168).The importance
of plasticization by water becomes more evident when one considers the hygroof water plasticiscopic nature of most dehydrated foods. A typical manifestation
zation is the loss of crunchiness or crispness in snack foods and breakfast cereals
(161,169,170). It is important to note that the loss of crispness in the corn Hour
* It
is interesting to note here that Matveev et al. (165) calculated Tg of proteins from the amino
acid composition and reported good agreement with the experimental results.
lastic
Linear
extrudates and in the corn cakes occurred within the glassy state, meaning
that
the knowledge of Tg alone may not be sufficient to predict crispness (142,171).
in thelowWater can also exertanantiplasticizingeffectparticularly
moisture or low-a, region by causing an increase in puncture strength, modulus,
and brittleness (167,172). Moreover, addition of low molecular weight diluents
other than water (e.g., fructose, glycerol) to glassy polymers (e.g., amylopectin,
starch) on the one hand lowers Tg but at the same time exerts an antiplasticizing
effect on the mechanical properties of the polymer ( 172- 175).
IV. SUMMARYANDFUTURETRENDS
Evaluation of food quality parameters has been traditionally performed via sensory panel tests and/or destructive mechanical tests suchas TPA and the uniaxial
compression and tension. Such methodsstill remain popular as they relate reasonablywell to the consumer perception of macroscopic food quality. However,
increasing consumer demand for consistently high-quality products and the comto obtain fundamental information
petitive marketplace have led the food industry
to improve quality assurance and/or control operations. Linear viscoelastic methods arewidely used for obtaining information regarding structural organization
of
the food materials. Since food texture
is largely a manifestation of microstructural
organization, viscoelastic methods provide the basic information needed
to understand the factors that influence quality.
The transient methods such as stress relaxation and creep have been the
most commonly used viscoelastic methods. By appropriate mechanical models,
transient test data can be used to obtain structural information and determine the
viscoelastic properties of the materials. The theory of dynamic viscoelastic tests
has long been known (176) and has been used in the polymer industry (46,177).
However, the application of dynamic tests for studying food materials have been
to eitherlack of properinstrumentation
delayed until relativelyrecentlydue
and/or highcost. The advances in computerandinstrumentationtechnology
have led to improved rheometer design at an affordable cost. This has promoted
widespread use of dynamic rheological testing of foods. While the transient tests
are relatively easier to perform, they may take a long time, whereas the dynamic
a
tests allow us to obtain viscoelastic parameters fairly accurately and within
shorttime (at frequenciesabove 1 Hz). In addition,theability
of dynamic
rheometers to perform tests at a wide range of frequencies provides information
not easily obtainable by other methods.
In the case of high-speed processing
of foods, for example, viscoelastic characterization at a high frequency is very
critical.
Determining the fundamental structural characteristics of foods is still the
main focus of linear viscoelastic tests. Nevertheless, new research efforts have
324
Gunasekaran
Ak and
been afoot in correlating transientand dynamic rheological parametersto various

sensory quality attributes and some relevant functional properties. We expect
in such attemptswill come only if the methsuch efforts to continue. But success
ods and the properties are well rooted in the fundamental structural aspects of
both the food materials and the targeted properties. In this regard, we hope that
more researchers will use the linear viscoelastic data to elucidate the structurefunction relationshipsof foods rather than to report trends and/or to obtain simple
statistical correlations with some enduse properties.
Studies on glass transition phenomena in amorphous foods are expected to
increase at a faster rate since techniques otherthan DSC, such as DMA, DMTA,
NMR, and ESR, are being introduced and used more and more. Complementary
data from allof these techniqueswould enable researchers to understand changes
taking place at and around the glass transition temperature in wide range of systems, from simple solutions to complex frozen foods. The plasticizing and antiplasticizing effectsof water and other diluents obviously deservemore investigation. Also, more work is expected to be done on the changes that occur in the
glassy state itself and how these changes affect the quality or textural attributes
of foods.
Although linear viscoelastic methods are useful, they suffer from the fact
that they employ very small strains to be relevant to study material properties
during real processes that often employ large and rapid deformations (e&, food
mixing and extrusion). Novel rheometers have been developed to address these
issues, which are nonlinear (28,178). These rheometers have expanded the ability
of rheologists to obtain information heretofore unavailable but critical for successful food process design and/or product development (179). We expect that
further research in the nonlinear viscoelasticity in conjunction with the results
obtained via linear viscoelastic tests will bring food rheologists closer to characterizing food material properties and quality attributes more completely than is
now possible.
Dedication: The authors dedicate this Chapter to the memory of thousands of

victims of the earthquakes in Turkey on August 17 and November 12, 1999.
REFERENCES
HR Moskowitz. Food quality: conceptual and sensory aspects. Food Qual Pref
6:
157-162,1995.
2. AV Cardello.Foodquality: relativity, context and consumerexpectations.Food
Qual Pref 6: 163-170, 1995.
3. H Lawless. Dimensions of sensory quality: a critique. Food Qual Pref 6:191-199,
1995.
1.
stic
Linear
325
4. MC Bourne. Food Texture and Viscosity: Concept and Measurement. New York:
Academic Press, Inc., 1982.
5. HT Lawless, H Heymann. Sensory Evaluation of Food: Principles and Practices.
New York: Chapman & Hall, 1998.
6. AS Szczesniak. Sensory texture profiling: historical and scientific perspective. Food
Techno152(8):54-57,1998.
7. ZM Vickers. Evaluation of crispiness. In: JMV Blanshard, JR Mitchell, eds. Food
Structure-Its Creation and Evaluation. London: Buttenvorths, 1988, pp 433-448.
8. JG Kapsalis, ed. Objective Methods in Food Quality Assessment. Boca Raton, FL:
CRC Press, Inc., 1986.
9. HR Moskowitz, ed. Food Texture: Instrumental and Sensory Measurement. New
York: Marcel Dekker, Inc., 1987.
as a reflection of microstructure.
10. M Langton, A Astrom, A-M Hermansson. Texture
Food Qual Pref 7(3/4):185-191, 1996.
1 I . JF Meullenet, BG Lyon, JA Carpenter, CE Lyon. Relationship between sensory
and instrument texture profile attributes. J Sensory Stud 13:77-93, 1998.
12. RL Jackman, DW Stanley. Creep behavior of tomato pericarp tissue as influenced
by ambienttemperatureripeningandchilledstorage.
J TextStud 26537-552,
1995.
13. RE Pitt, HL Chen. Time-dependent aspectsof the strength and rheology of vegetative tissue. Trans ASAE 26:1275-1280, 1983.
14. N N Mohsenin. Physical Propertiesof Plant and Animal Materials: Structure, Physical Characteristics and Mechanical Properties. 2nd ed. New York: Gordon Breach
Science Publishers, 1986.
15. RE Pitt. Viscoelastic properties of fruits and vegetables. In:
MA Rao, JF Steffe,
eds. Viscoelastic Properties of Foods. New York: Elsevier Applied Science, 1992,
pp 49-76.
16. SB Ross-Murphy. Small deformation measurements. In: JMV Blanshard, JR Mitchell, eds. Food Structure-Its Creation and Evaluation. London: Butterworths, 1988,
pp 387-400.
17. CW Macosko. Rheology: Principles, Measurements, and Applications. New York:
VCHPublishers,1993,pp1-62.
18. J Ferguson, Z Kemblowski. Applied Fluid Rheology. New York: Elsevier Applied
Science,1991,pp1-45.
19. M Reiner. The Deborah number. Physics Today 17(1):62, 1964.
20. HA Barnes, JF Hutton,KWalters. An IntroductiontoRheology.Amsterdam:
Elsevier Science Publishers BV, 1989, pp 1-54.
21. R Subramanian,S Gunasekaran. Small amplitude oscillatory shear studies on Mozzarella cheese. Part 11. Relaxation spectrum. J Text Stud 28:643-656, 1997.
22. JL Kokini, C-T Ho, MV Kanve. Food Extrusion Science and Technology. New
York: Marcel Dekker, 1992.
23. CA Tolman, NB Jackson. The coupling of physical and chemical effects. In: WT
Lippincott, ed. Essays in Physical Chemistry: A Sourcebook
for Physical Chemistry
Teachers. Washington, DC: American Chemical Society, 1988, p 22.
24. G Schramm. A Practical Approach to Rheology and Rheometry. Karlsruhe, Germany: Haake, GmbH, 1994.
326
Gunasekaran
Ak and
25. FM Ernsberger.Elasticproperties of glasses.In: DR Uhlmann,NJKreidl,eds.

Glass: ScienceandTechnology.Vol. 5. New York:AcademicPress,Inc.,1980,
pp 1-17.
26. RC Plumb. Antique windowpanes and the flow of supercooled liquids. J Chem Ed
66( 12):994-996, 1989.
27.EDZanotto.Docathedralglassesflow?
Am J Phys66(5):392-395.1998.
28.RSLakes.ViscoelasticSolids.BocaRaton,
FL: CRCPress,1998.
29. JJ Aklonis, WJ MacKnight, M Shen. Introduction to Polymer Viscoelasticity. 2nd
ed. NewYork:Wiley-Interscience,1983,pp1-87.
30. S Purkayastha,MPeleg.Comparisonbetweenprojectedmechanicalequilibrium
conditions of selected food materials
in stress relaxation and creep. J Text Stud
17:433-444.1986.
31. MPeleg.MDNormand.Comparisonoftwomethodsforstressrelaxationdata
presentation of solid foods. Rheol Acta 22: 108-1 13, 1983.
32. S Purkayastha, M Peleg, MD Normand. Presentation ofthe creep curves of solid
biological materialsby a simplified mathematical version
of the generalized KelvinVoigt model. Rheolog Acta 2336-563, 1984.
33. NW Tschoegl. The Phenomenological Theory of Linear Viscoelastic Behavior: An
Introduction. New York: Springer-Verlag, 1989.
34.ASNavarro, MN Martino, NE Zaritzky.Correlationbetweentransientrotational
viscometry and a dynamic oscillatory test for viscoelastic starch based systems, J
Text Stud 28(4):365-385, 1997.
35. BG Diener, DR Heldman. Methods of determining rheological properties of butter.
TransASAE11:444-447, 1968.
36. NG McCrum, CP Buckley, CB Bucknall. Principles of Polymer Engineering. New
York: Oxford Science Publications, Oxford University Press, 1992, pp 1 17-170.
37. H Rohm, K-H Weidinger. Rheological behaviour of butter at small deformations.
J Text Stud 24:157-172, 1993.
38. DWeipert.Determiningrheologicalproperties
of cerealproductsusingdynamic
mechanical analysis in compression mode. Cereal Foods World 42(3): 132- 137,
1997.
39. JD Ferry. Viscoelastic Properties of Polymers. 3rd ed.
NewYork:JohnWiley &
Sons,Inc.,1980.
40. JG Ceder,DLOutcalt.Calculus.Boston:Allyn
& Bacon,Inc.,1977.
41. JM Dealy. Official nomenclature for material functions describing the response of
a viscoelastic fluid to various shearing and extensional deformations.
J Rheol39( I):
253-265,1995.
42. MMAk, S Gunasekaran.Dynamicrheologicalproperties
of Mozzarellacheese
during refrigerated storage. J Food Sci 61(3):566-568, 584, 1996.
43. R Subramanian, S Gunasekaran. Small amplitude oscillatory shear studies on Mozzarella cheese. Part I. Region of linear viscoelasticity. J Text Stud 28:633-642,
1997.
44. JMV Blanshard, PJ Lillford, eds. The Glassy State in Foods. Loughborough, UK:
NottinghamUniversityPress,1993.
45. HH Winter. Analysis of dynamic mechanical data: inversion into a relaxation time
spectrum and consistency check. J Non-Newtonian Fluid Mech 68:225-239, 1997.
327
46. LE Nielsen, RF Landel. Mechanical Properties of Polymers and Composites. 2nd

ed. New York: Marcel Dekker, Inc., 1994.
47. RW Whorlow. Rheological Techniques. Chichester, England: Ellis Harwood Ltd..
1980.
48. GV Gordon, MT Show. Computer Programs for Rheologists. Cincinnati: Hanser/
Gardner Publications, 1994, pp 72-95.
49. DHS Ramkumar, JM Wiest. Molecular weight distributions from linear viscoelastic
measurements. Rheol Acta 35:356-363, 1996.
a study
50. c Carrot, J Guillet. From dynamic moduli to molecular weight distribution:
of various polydisperse linear polymers. J Rheol 41(5):1203-1220, 1997.
51. JM Dealy, KF Wissbrun. Melt Rheology and Its Role in Plastic Processing: Theory
and Applications. New York: Van Nostrand Reinhold, 1989.
52. V Evageliou, S Alevisopoulos, S Kasapis. Application of stress-controlled analysis
to the development of low fat spreads. J Text Stud 28:319-335, 1997.
53. H Rohm. Rheological behavior of butter at large deformations. J Text Stud 24:
139-155,1993.
54. IS Chronakis. S Kasapis. A rheological study on the application of carbohydrateprotein incompatibility to the development of low fat commercial spreads. Carbohydr Polym 28:367-373, 1995.
55. N Sakurai, DJ Nevins. Evaluation of stress-relaxation in fruit tissue. Hort Techno1
2398-402, 1992.
56. N Sakurai, OJ Nevins. Changes in physical properties and cell wall polysaccharide
of tomato (Lycopersicon eseulentum) pericarp tissue. Physiol Plast 89:68 1-686,
1993.
57. JP Mittal, N N Moshenin. MG Sharma. Rheological characterizationof apple cortex.
J Text Stud 18:65-93, 1987.
58. MA Rao. Viscoelastic properties of fluid and semisolid foods. In: MR Okos, ed.
Physical and Chemical Properties of Foods. St. Joseph, MI: American Society of
Agricultural Engineers, 1986, pp 14-34.
59. VT Karathanos, AE Kostaropoulos, GD Saravacos. Viscoelastic properties of raisins. J Food Eng 23:481-490, 1994.
60. P Lewicki, W Wolf. Rheological properties of raisins: Part 1. Compression tests.
J Food Eng 24321-338. 1995.
61. OH Campanella, M Peleg. On food compressionby soft machines. J Text Stud 19:
39-50,1988.
62. S Purkayastha, M Peleg, EA Johnson, MD Normand. A computer aided characterization of the compressive creep behavior of potato and Cheddar cheese. J Food
Sci 50:45-50, 55, 1985.
63. L Ma, MA Drake, GV Barbosa-Canovas, BG Swanson. Viscoelastic properties of
reduced-fat and full-fat Cheddar cheeses. J Food Sci 61(4):821-823, 1996.
CR Daubert.Rheologicalandsensorypropertiesof
64. MA Drake.VDTruong,
reduced-fat processed cheeses containing lecithin.
J Food Sci 64(4):744-747, 1999.
65. M-l Kuo, Y-C Wang,S Gunasekaran. A viscoelasticity index for cheese meltability
evaluation. J Dairy Sci 83(3):1-6, 1999.
66. MG Lynch, DM Mulvihill. The influence of caseins on the rheology of 1-carrageenan gels. Food Hydrocolloids 8 3 17-329, 1994.
328
Ak and Gunasekaran
67. Y-C Wang, K Muthukumarappan, MM Ak, S Gunasekaran. A devicefor evaluating

meltlflow characteristics of cheeses. J Text Stud 29:43-55, 1998.
68. RC Hoseney, JO Smewing. Instrumental measurement of stickiness of doughs and
other foods. J Text Stud 30:123-136, 1999.
69. BJ Dobraszczyk. The rheological basis of dough stickiness.
J Text Stud 28: 139162,1997.
70. P Dejmek. Dynamic rheology of rennet curd.
J Dairy Sci 70: 1325-1330, 1987.
71.PZoon,T
vanVliet,PWalstra.Rheologicalpropertiesofrennet-inducedskim
milk gels. 1. Introduction. Netherlands Milk Dairy J 42:249-269, 1988.
72.MPaulsson,PDejmek,TvanVliet.Rheologicalpropertiesofheat-induced
plactoglobulin gels. J Dairy Sci 73:45-53, 1990.
73. P Zoon,TvanVliet,PWalstra.Rheologicalproperties
of rennet-inducedskim
milk gels. 2. The effect of temperature. Neth Milk Dairy J 42:271-294, 1988.
74.LBohlin, P-0 Hegg,HLjusberg-Wahren.Viscoelasticpropertiesofcoagulating
milk. J Dairy Sci 67:729-734, 1984.
75. K Nakamura. R Niki. Rheological properties of casein micelle gels: the influence
of calcium concentration on gelation induced
by rennet. Biorheology 30(3-4):207216,1993.
76. DL van Hekken, ED Strange. Rheological properties and microstructure of dephosphorylated whole casein rennet gels. J Dairy Sci 77:907-916, 1994.
77. TP Guinee, DJ OCallaghan, PD Pudja, N OBrien. Rennet coagulation properties
of retentates obtained by ultrafiltration of skim milks heated to different temperatures. Int Dairy J 6581-596, 1996.
78. A Skriver. Characterization of stirred yoghurt by rheology, microscopy and sensory
analysis. PhD dissertation, Institute for Dairy Research, The Royal Veterinary and
Agricultural University, Denmark, 1995.
79. SB Ross-Murphy. Structure-property relationships in food biopolymer gels and
solutions. J Rheol 39(6):1451-1463, 1995.
80. A Skriver, H Roemer, KB Qvist. Rheological characterization of stirred yoghurt:
viscometry. J Text Stud 24: 185-198, 1993.
8 I . E Ronnegird, P Dejmek. Development and breakdown of structurei n yoghurt studied by oscillatory rheological measurements. Lait 73:371-379, 1993.
82. M Arshad, M Paulsson, P Dejmek. Rheology of buildup. breakdown, and rebodying
of acid casein gels. J Dairy Sci 76:3310-3316, 1993.
83. JA Lucey.TvanVliet, K Grolle,TGeurts,PWalstra.Properties
Of acid Casein
gelsmadebyacidificationwithglucono-6-lactone.
1. Rheological properties. Int
Dairy J 7:381-388, 1997.
84. A Salvador, SM Fiszman. Textural characteristics and dynamic oscillatory rheology
of maturation of milk gelatin gels with low acidity.
J Dairy Sci 8 I(6): 1525- 1531,
1998.
85. AS Navarro,CFerrero,
NE Zaritzky.Rheologicalcharacterizationof
Duke
de Lethe" bydynamicandsteadyshearmeasurements.
J Text Stud 30:43-58,
1999.
86. BH Ozer, RK Robinson,ASGrandison, AE Bell.Comparison of techniques for
measuring the rheological properties of labneh (concentrated yogurt). Int J Dairy
Techno1 50(4):129-134,1997.
stic
Linear
329
87. MRuegg,PEberhard,LMPopplewell,MPeleg.Meltingpropertiesofcheese.
Bull. 268, Int Dairy Fed, Brussels, Belgium, 1991, pp 36-43.
88. J Park, JR Rosenau, M Peleg. Comparisonof four procedures of cheese meltability
evaluation. J Food Sci 49:1158-1161, 1170, 1984.
89. Z Ustunol, K Kawachi, J Steffe. Arnott test correlates with dynamic rheological
properties for determining Cheddar cheese meltability. J Food Sci 59(5):970-971,
1994.
90. JS Mounsey, ED ORiordan. Empirical and dynamic rheological data correlation
to characterize melt characteristics of imitation cheese.J Food Sci 64(4):701-703,
1999.
91. M Sutheerawattananonda, ED Bastian. Monitoring process cheese meltability using
dynamic rheometry. J Text Stud 29:169-183, 1998.
92. AS Yoshimura, RK Prudhomme. Wall slip effects on dynamic oscillatory measurements. J Rheol 32(6):575-584, 1988.
93. EJ Nolan, VH Holsinger, JJ Shieh. Dynamic rheological properties of natural and
imitation mozzarella cheese. J Text Stud 20: 179- 189, 1989.
94. MH Tunick, EJ Nolan,JJShieh,
JJ Basch, MP Thompson,BEMaleef,VH
Holsinger. Cheddar and Cheshire cheese rheology. J Dairy Sci 73(7):1671-1675,
1990.
95. MH Tunick, KL Mackey, JJ Shieh, PW Smith, P Cooke, EL Malin. Rheology and
microstructure of low-fat Mozzarella cheese. Int Dairy J 3:649-662, 1993.
96. M Rosenberg, Z Wang, SL Chuang, CF Shoemaker. Viscoelastic property changes
in Cheddar cheese during ripening. J Food Sci 60(3):640-644, 1995.
97. MM Ak, S Gunasekaran. Evaluating rheological properties of Mozzarella cheese
by the squeezing flow method. J Text Stud 26:695-71 I , 1995.
98. OHCampanella, LM Popplewell,JRRosenau,MPeleg.Elongationalviscosity
measurements of melting American process cheese. J Food Sci 52(5): 1249-1251,
1987.
99. EM Casiraghi, EB Bagley, DD Christianson. Behavior of Mozzarella, Cheddar and
Processed cheese spread in lubricated and bonded uniaxial compression.
J Text
Stud16:281-301,1985.
100. HH Winter, F Chambon. Analysis of linear viscoelasticity
of
a crosslinking
polymer
at the gel point. J Rheol 30(2):367-382, 1986.
101 F Chambon, HH Winter. Linear viscoelasticity at the gel point of
a crosslinking
PDMS with imbalanced stoichiometry. J Rheol 31(8):683-697, 1987.
102. HH Winter. Can the gel point of
a cross-linking polymer be detected by the GG crossover? Polymer Eng and Sci 27(22):1698-1702, 1987.
103. YG Lin, DT Mallin, JCW Chien, HH Winter. Dynamic mechanical measurement
of crystallization-induced gelation in thermoplastic elastomeric poly(propy1ene).
Macromolecules24:850-854,1991.
104. HH Winter. Polymer gels, materials that combine liquid and solid properties. MRS
Bull XVI (8):44-48, 1991.
105. SB Ross-Murphy. Rheological characterisation of gels.
J Text Stud 26:391-400,
1995.
106. C-YM Tung, PJ Dynes. Gelation in thermosetting systems. J Appl Polym Sci 27:
569-574.1982.
330
Gunasekaran
Ak and
107.JALopesdaSilva,MPGonqalves.Rheologicalstudyintotheagingprocess
of highmethoxylpectinlsucroseaqueousgels.CarbohydrPolym24:23.5-245,
1994.
108. CMichon, G Cuvelier, B Launay,AParker.Sol-geltransition
of 1-carrageenan
andgelatinsystems:dynamicvisco-elasticcharacterization.
In: EDickinson.D
Lorient, eds. Food Macromolecules and Colloids. Cambridge: The Royal Society
of Chemistry, 1995, pp 462-47 I.
109.AELabropoulos,S-HHsu.Viscoelasticbehaviorofwheyproteinisolatesatthe
sol-gel transition point. J Food Sci 61( 1):65-68. 1996.
1 10. KLarsson, SE Friberg.eds.FoodEmulsions.
New York:MarcelDekker,Inc..
1990.
1 1 1 . DG Dalgleish. Adsorption of protein and the stability of emulsions. Trends Food
SciTechno1 8:l-6, 1997.
I 12.EDickinson.MGolding.Rheologyofsodiumcaseinatestabilizedoil-in-water
emulsions. J Coll Interface Sci 19 1 : 166- 176, 1997.
I 13. J Muiioz, P Sherman. Dynamic viscoelastic properties of some commercial salad
dressings. J Text Stud 2 1:4 1 1-426, 1990.
I 14. C Gallegos, M Berjano, L Choplin. Linear viscoelastic behavior of commercial and
model mayonnaise. J Rheol 36(3):465-478, 1992.
I 1.5. L Ma, GV Barbosa-Cinovas. Rheological characterization of mayonnaise. Part 11:
flow and viscoelastic properties at different oil and xanthan gum concentrations. J
Food Eng 25:409-425, 1995.
116. R Jost,FDannenberg, J Rosset.Heat-setgelsbased
on oillwateremulsions:an
application of whey protein functionality. Food Microstructure 8:23-28, 1989.
117. E Dickinson, Y Yamamoto. Effect of lecithin on the viscoelastic properties of plactoglobulin-stabilized emulsion gels. Food Hydrocoll 10(3):301-307, 1996.
1 18. JMFranco, A Guerrero,CGallegos.Rheologyandprocessingofsaladdressing
emulsions. Rheol Acta 34:s 13-524, 199.5.
119. RK Richardson, ER Morris, SB Ross-Murphy, LJ Taylor, ICM Dea. Characterization of the perceived texture of thickened systems by dynamic viscosity measurements. Food Hydrocoll 3:17.5-191, 1989.
120. ZV Baines, ER Morris.Flavour/tasteperceptioninthickenedsystems:theeffect
of guar gum above and below C*. Food Hydrocoll 1:197-205, 1987.
121. ER Morris. Rheology and organoleptic properties of food hydrocolloids. In: K Nishinari, E Doi. eds. Food Hydrocolloids. New York: Plenum Press, 1993, pp 201210.
122. MA Hill, JR Mitchell, PA Sherman.
The relationship between the rheological and
sensory properties of a lemon pie filling. J Text Stud 26:4.57-470, 199.5.
123. RM Pangborn, IM Trabue, AS Szczesniak. Effect of hydrocolloidson oral viscosity
and basic taste intensities. J Text Stud 4:224-241, 1973.
124. ZV Baines, ER Morris. Suppression of perceived flavour and taste by food hydrocolloids. In: RDBee,PRichmond, J Mingens,eds.FoodColloids.Cambridge,
England: Royal Society of Chemistry, 1989, pp 184-192.
125. ZV Baines, ER Morris. Effect of polysaccharide thickeners on organoleptic attributes. In: GO Philipps, DJ Wedlock, PA Williams, eds. Gums and Stabilisers for
the Food Industry, vol. 4. Oxford, England: IRL Press, 1988, pp 193-201.
331
126. EJ Win&&,BWolf.EByekwaso.Structure-rheology-texturerelationshipsfor
colloidal food systems. In: E Dickinson, B Bergenstahl, eds. Food Colloids: Proteins, Lipids and Polysaccharides. Cambridge: The Royal Society of Chemistry,
1987,pp12-13.
127. BBRfJnn, G Hyldig, L Wienberg, KB Qvist, AM Laustsen.Predictingsensory
properties from rheological measurements
of low-fat spreads. Food Qual Preference
9(4):187-196,1998.
128. YH Roes, M Karel, JL Kokini. Glass transition in low moisture and frozen foods:
effects on shelf life and quality. Food Technol 50(11):95-108, 1996.
129. M Peleg. Description of mechanical changes in foods at their glass transition region.
In:GVBarbosa-Canovas, J Welti-Chanes,eds.FoodPreservation
by Moisture
Control: Fundamentals and Applications. Lancaster, PA: Technomic Press. 1994,
pp 659-67 I .
130. M LcMeste, VT Huang, J Panama, G Anderson, R Lentz. Glass transition of bread.
Cereal Foods World 37(3):264-267, 1992.
131. HD Goff. Low-temperature stability and the glassy statein frozen foods. Food Res
Int 25317-325, 1992.
132. HD Goff. Measuring and interpreting theglass transition in frozen foods and model
systems. Food Res Int 27: 187-189. 1994.
133. YH Roos. Glass transition-related physicochemical changes in foods. Food Technol
49(10):97-101,1995.
glass transition-dependence of the glass transition
134. L Slade, H Levine. Water and the
on composition and chemical structure: special implications for flour functionality
in cookie baking. J Food Eng 24:431-509, 1995.
135. KA Nelson, TP Labuza. Water activity and food polymer science: implications of
state on Arrhenius and WLF models in predicting shelf life. J Food Eng 22:271289,1994.
136. TR Noel, SG Ring, MA Whittam. Glass transitions in low-moisture foods. Trends
FoodSciTechnol1:62-67,1990.
137. L Slade, H Levine. The glassy state phenomenon in food molecules. In: JMV Blanshard, PJ Lillford,eds.TheGlassyStateinFoods.Loughborough:Nottingham
University Press, 1993, pp 35-101.
138. DS Reid, W Kerr, J Hsu. The glass transition in the freezing process. J Food Eng
221483-494,1994.
139. BR Bhandari, T Howes. Implication of glass transition for the drying and stability
of dried foods. J Food Eng 40:7 1-79, 1999.
140. H Watanabe, CQ Tang, T Suzuki, T Mihori. Fracture stress of fish meat and the
glass transition. J Food Eng 29:317-327, 1996.
141. D Torreggiani, E Forni,I Guercilena, A Maestrelli,G Bertolo, GP Archer,CJ Kennedy, S Bone, G Blond, E Contreras-Lopez, D Champion. Modifications of glass
transition temperature through carbohydrates additions effect upon colour and anthocyanin pigment stability in frozen strawberry juices. Food Res Int 32:44 1-446,
1999.
142 Y Li, KM Kloeppel, F Hsieh. Texture of glassy corn cakesas function of moisture
content. J Food Sci 63(5):869-872, 1998.
in relation to
I43 WLKerr, MH Lim, DS Reid, H Chen. Chemical reaction kinetics
332
Gunasekaran
Ak and
glass transition temperatures in frozen food polymer solutions.
J Sci Food Agric

61:51-56,1993.
144. Y-J Wang, J Jane. Correlation betweenglass transition temperature and starch retrogradation in the presence of sugars and maltodextrins. Cereal Chem 71(6):527531, 1994.
145. Y Shimada, Y Roos, M Karel. Oxidation of methyl linoleate encapsulated in amorphous lactose-based food model. J Agric Food Chem 39(4):637-641, 1991.
146. R Karmas, MP Buera, M Karel. Effect of glass transition on rates of nonenzymatic
browning in food systems. J Agric Food Chem 40(5):873-879, 1992.
147. J Perez. Theories of liquid-glass transition. J FoodEng22:89- 114, 1994.
148. AM Cocero, JL Kokini. The study of the glass transition of glutenin using small
amplitude oscillatory rheological measurements and differential scanning calorimetry. J Rheol 35(2):257-270, 1991.
149.MTKalichevsky, EM Jaroszkiewicz, S Ablett,JMVBlanshard, PJ Lillford.The
glass transition of amylopectin measured by DSC, DMTA and NMR. Carbohydr
Polym18:77-88,1992.
150. MAHemminga,MJGWRoozen,PWalstra.Molecularmotionsandtheglassy
state. In: JMV Blanshard, PJ Lillford, eds. The Glassy State in Foods. Loughborough:NottinghamUniversityPress,1993,pp157-171.
15 I . MT Kalichevsky, JMV Blanshard, RDL Marsh. Applications of mechanical spectroscopy to the study of glassy biopolymers and related systems. In: JMV Blanshard,PJLillford,eds.TheGlassyState
in Foods.Loughborough:Nottingham
University Press, 1993, pp 133-156.
152. A Nikolaidis, TP Labuza. Glass transition state diagrams of a baked cracker and
its relationship to gluten. J Food Sci 61(4):803-806, 1996.
153. AMizuno,MMitsuiki,MMotoki.Glasstransitiontemperatureofcasein
as affected by transglutaminase. J Food Sci 64(5):796-799, 1999.
154.GAllen. A history of theglassystate.In:JMVBlanshard,
PJ Lillford,eds.The
Glassy State in Foods. Loughborough: Nottingham University Press, 1993, pp
112.
155. M Peleg. Mapping the stiffness-temperaturemoisture relationship of solid biomaterials at and around their glass transition. Rheol Acta 32575-580. 1993.
156.MPeleg.Glasstransitionsandthephysicalstabilityoffoodpowders.In:JMV
Blanshard, PJ Lillford, eds. The Glassy State
in Foods. Loughborough: Nottingham
Univ. Press, 1993, pp 435-45 I .
157. M Peleg. A note on the tan
6 (T) peak as a glass transition indicator in biosolids.
Rheol Acta 34:215-220, 1995.
158. M Peleg.Onmodelingchangesinfoodandbiosolids
at andaroundtheir glass
transition temperature range. Crit Rev Food Sci Nulr 36(1&2):49-67, 1996.
159.MLWilliams,
RF Landel, JD Ferry.Thetemperaturedependenceofrelaxation
lnechanislns in amorphous polymers and other glass-forming liquids. J Am Chem
Soc 773701-3707, 1955.
Of the stiffness160. M Peleg. Mathematical characterization and graphical presentation
temperature-nloisture relationship of gliadin. Biotechnol Prog I0:652-654. 1994.
161.MHarris,MPeleg.Patterns
of texturalchanges i n brittlecellularcerealfoods
caused by moisture sorption. Cereal Chem 73(2):225-231, 1996.
333
162. M Peleg. A mathematical model of crunchiness/crispness loss in breakfast cereals.

J Text Stud 25:403-410, 1994.
163. R Hagen, L Salmen, H Lavebratt, B Stenberg. Comparisonof dynamic-mechanical
measurementsandTgdeterminationswith2differentinstruments.PolymTest
13(2):113-128,1994.
glass
164. MT Kalichevsky, JMV Blanshard. The effect of fructose and water on the
transition of amylopectin. Carbohydr Polym 20: 107- 1 13, 1993.
165. Y1 Matveev, VY Grinberg, IV Sochava, VB Tolstoguzov. Glass transition temperature of proteins. Calculation based on the additive contribution method and experimental data. Food Hydrocoll 1 l(2): 125-133, 1997.
166. LM Hallberg, P Chinachoti. Dynamic mechanical analysis for glass transitions in
long shelf-life bread. J Food Sci 57(5):1201-1204, 1229, 1992.
167. N Gontard, S Guilbert, J-L Cuq. Water and glycerol as plasticizers affect mechanical and water vapor barrier properties of an edible wheat gluten film. J Food Sci
58(1):206-21 I , 1993.
168. YH Roos. Water activity and physical state effects on amorphous food stability. J
FoodProcPres16:433-447,
1993.
169. F Sauvageot, G Blond. Effect of water activity on crispness of breakfast cereals.
J Text Stud 22:423-442, 1991.
170. M Wollny, M Peleg. A model
of moisture induced plasticization of crunchy snacks
on Fermis distribution function. J Sci Food Agr 64:467-473, 1994.
171. G Kaletunc, KJ Breslauer. Glass transitionofextrudates:relationshipwithprocessing-inducedfragmentationandend-productattributes.CerealChem70(5):
548-552, 1993.
172. CC Seow, PB Cheah, YP Chang. Antiplasticization by water
in reduced-moisture
food systems. J Food Sci 64(4):576-58 I , 1999.
173. M Peleg. Mathematical characterization of the plasticizing and antiplasticizing effects of fructose on amylopectin. Cereal Chem 73(6):712-715, 1996.
174. D Lourdin, H Bizot, P Colonna. Antiplasticization
in starch-glycerol films? J
Appl Polym Sci 63: 1047-1053, 1997.
17.5 S Gaudin, D Lourdin, D Le Botlan,JL Ilari, P Colonna. Plasticisation and mobility
i n starch-sorbitol films. J Cereal Sci 29(3):273-284, 1999.
176. DR Bland. The Theory of Linear Viscoelasticity.New York, NY: Pergamon Press,
1960.
177. HC Robin. Computer Aided Analysis of Stress/Strain Response of High Polymers.
Lancaster, PA: Technomic Publishing, 1989.
178. AJ Giacomin, T Smurkas, JM Dealy. A novel sliding plate rheometer for molten
plastics. Polymer Eng Sci 35:499-504,
1989.
179. S Tariq, AJ Giacomin, S Gunasekaran. Nonlinear viscoelasticity of cheese.J Biorheol35(3):171-191,1998.
Biosensors in Food
Quality Evaluation
Sudhir S. Deshpande
Signature Bioscience, Inc., Burlingame, California
1.
INTRODUCTION
Quantitative determination of the composition and propertiesof both raw materials and processed food products is important from the viewpoint of quality and
safety of the food. Quality depends on color, flavor, aroma, texture, nutritional
attributes, and microbial content. Food safety relatesto the content of pathogenic
microorganisms and toxic compounds, including microbially produced toxins,
pesticides, and naturally occurring toxic and antinutritional compounds, as well
as compounds produced by processing. Food analysis thus presents several chalof foods available,
lenges, in large part because each analyte in the wide range
from fluid products such as milk and fruit juices to processed meats, is complex
and nonhomogeneous. Quality control testingin the food industry has been based
upon traditional analytical methods that require hours, if not days, to complete.
In recent years, the food industry is increasingly adopting food safety and
quality management systems that are more proactive and preventive than those
to rely on end-product testing and visual
used in the past, which have tended
inspection. The regulatory agencies in many countries are promoting one such
management tool, Hazard Analysis Critical Control Point (HACCP), as a way
to achieve a safer food supply and also as a basis for international harmonization
of trading standards. For HACCP to be effective and successful, rapid methods
for monitoring and verifying performance need to be available. This, in turn, has
provided an impetus for developing newer and improved detection methods and
technologies that are cheaper, better, and faster.
Until recently, rapid quality-monitoring methods for on- and near-line use
were considered not to be very accurate or very reliable. The consensus was that
335
336
Deshpande
the fastest resultis not always the best. The

199Os, however, have seen the procession of traditional laboratory analyses to the factory production floors. Measurement of constituents such as fat, protein, carbohydrates, moisture, and solids is
being incorporated into quick, convenient instruments designed for use in close
physicalorfunctionalproximitytotheprocess.Withtimelyfeedback,better
process decisions are made, helping boost production volume, reduce inventory
and product in hold, and free up laboratory personnel. Sometimes, incorporation
of sophisticated instrumentation on-line requires high initial investments.But the
tremendous cost savings that rapid methods can help realize, in terms of wasted
product and time lost pending analytical results, should not be taken lightly. Although such investment depends on the product and production volume, often it
pays off within a matter of months. Along with costs, several points need to be
considered in selecting on-line sensors for monitoring food process and quality
control. Some of these are summarized in Table I .
In modern food-processing plants, computer control has already become
an important part of factory operation. As yet unavailable, or scarcely used, are
Table 1 Factors to Consider in Using On-line Sensor Technologies for Food Process
and Quality Control

Factor
How does the method's accuracy compare to standard methods?
What is the percent recovery?
Does the method identify and measure parameters of
interest? What
Specificity
steps are taken to ensure a high degree of specificity'?
Precision
Is the method's precision within batch. batch-to-batch, or day-to-day
variation'? What steps in the procedure contribute the greatest variability?
Ruggedness
Does the instrument operate consistently in tough conditions'?
Is it insensitive to fluctuations in pressure and temperature prevalent in
food plant settings?
Hardware
What are the initial and ongoing operational and maintenance costs?
How often does the instrument have to be calibrated. and what is
involved?
What are their costs and availability? Are there any disposal and exReagents and
posure concerns?
consumables
Are the
Reproducibility How reliable is the method from the standpoint of precision?
results of one set of measurements comparable to those from another set of measurements by the same instrument and those from
another similar instrument/method?
Does the method give results quickly enough'! What is the on-line
Speed
compatibility of the instrument with the existing and future computer control systems in the plant'?
Accuracy
Biosensors
337
sensors to simultaneously monitor the multiple parameters

of the production lines
and report the data to the computer. Furthermore, traditional monitoring usually
relies on surveillance of physical and/or chemical parameters of a process such
as the time and temperature of heating or pH. However, the issue of foodborne
pathogens has recently captured the attention and concern of the scientific community, academia, government agencies, and consumers. Much hasbeen written
about the number of cases, outbreaks, foods, and microorganisms involved and
as the
about the costs of foodborne diseases to the nations economy as well
potential negative impact of lack of food safety in the nations food supply ( 1 4). The culprits are the pathogenic microorganisms and their toxins and metabolites in our food supply. Thus, there is an increasing need for the introduction
of in-line microbiological analysis and the production of data relevant to product
safety and shelf life.
There is some controversy as to whether microbiological tests can be used
to monitor critical control points during food processing because of the length
of
time needed to generate results and the sampling strategy required for meaningful
results. Verification that the process is safe must involve microbiological testing,
but the results need not be generated in real time. However, considerable advantages may accrue if verification can be achieved quickly. The HACCP models
primary goal-on- or at-line monitoring control-has indeed driven a focus on
the role of microbiological methods for use in such programs.
a new, rapid typeof moniBiosensor technology can offer the food industry
toring and measuring device whose speed, sensitivity, stability, and ease
of use
exceed the current methodology. The ability
of the food microbiologist, for example, to monitor food productionon line with an organism probe, whose utilization is as simple and straightforward as that of a pH meter, is not yet a reality.
of biosensors may rapidly perform a variety
However, in the future an array
ofanalyticalprocedureswithhighreproducibility,specificity,andsensitivity.
Potential uses for biosensor technologies
in the food industry include, among
others, proximate analysis, nutritional labeling, pesticide residues, naturally occurring toxins and antinutrients, processing changes, microbial contamination,
enzymatic inactivation, and BOD (biological oxygen demand) of wastes.
In this chapter, the basic principles of the biosensor technology as relevant
to the food industry, a variety
of biosensors for analytical purposes, emerging
detection technologies in this field and their future in the food QA/QC laboratories are briefly described.
II. WHATAREBIOSENSORS?
A sensor canbe defined as a devicethat respondsto a physical stimulus producing

a response or signal that can be used for measurement, interpretation, or control.
It is usually constructed from three components: the detector, which recognizes
Deshpande
338
the physical stimulus; the transducer, which converts the stimulus into a suitable
signal; and the output system itself, which involves amplification, analysis, processing, and display that is usually electrical.
The term biosensor is now generally applied to those devices that employ a biologicaUbiochemica1 detection system. The reactive surface (e.g., the
biochemical or signal transducer) generates a measured response upon binding
with an analyte. The device and associated instrumentation (the physical transducer component) convert analyte-binding events into qualitative or quantitative
units of analyte concentration. Thus, the conversion of the analyte by the biochemical transducer into another chemical species and/or physical property is
sensed and converted into a measurable signalby the physical transducer. Immunosensors are essentially biosensors that use antibodies as signal transducers.
Molecular recognition of a specific sample analyte by the respective receptor is realized by the lock-and-key principle, which is the biocomponent or
the biocatalyst that gives the biosensor its specificity and selectivity ( 5 - 8 ) . It is
also involved in rapid recognition and detectionof specific molecules or chemical
reactions. This molecular recognition and detection mechanism, in general, creates a chemical change in or by the biological element proportional to the reaction, which can then be monitored by the physical transducer. The biocomponent
must, therefore, have intimate contact with a suitable transducing system. Enzymes, receptors, organelles, tissues, nucleic acids, antigens,
or antibodies, as
well as whole cells of bacterial, fungal, plant, or animal origin, can be used as
a biological detection system.
The physical transducers or measuring principles generally vary with the
type of biological detection system used in the biosensor. The transduction systems commonly used in biosensors are summarized in Table 2 . For example,
Table 2 CommonlyUsedBiosensorTransductionSystems
Transduction
system
Options
ElectrochemicalPotentiometry,amperometry,voltametry,ion-sensitiveelectrode
(ISE), ion-sensitive field effect transistor (ISFET), immuno-field effect transistor (IMFET)
Electrical
Surface
conductivity,
conductivity,
capacitance
Thermal
Calorimetry,
enzyme
thermistor
(heat
of
reaction
or
absorption;
TELISA or thermistor ELISA)
Paramagnetism
Magnetic
Optical
Fluorescence,
luminescence,
reflection,
absorption,
surface
plasmon
resonance, scattering, evanescent waves
PiezoelectricQCM.SAQ,
SH/APM, Lambwave,Lovewave
Biosensors
339
activity of an enzyme biocomponent can be monitored by measuring pH change,

production of oxygen or hydrogen peroxide in the chemical reaction, nicotinamide adenine dinucleotide (NADH) fluorescence or absorption, conductivity, or
changes in temperature. Similarly, reactions involving the antibody
as the biocomponent can be monitored via a surface plasmon resonator grating coupler,
interferometer piezoelectric device, fluorescence or enzyme-linked assays, elec(7). In contrast,
trodes, or absorption, fluorescence, or luminescence measurement
oxygen (0,)and carbon dioxide (CO,) electrodes are preferred transducers when
microbial cells or plant and animal tissues are usedas the main biological components (9).
The transducer can be viewed as the interpreter of the biosensor device. It
takes the energy impulse from the biocatalyst and translates
it into a usable signal
relative to the conditions of the measured analyte. Proper electronics then amplify
the signal and transferit into an output that can be interpreted by the user. Interference and distortion of the signal are minimized,as the transducer is usually adjacent to the biocatalyst. The interface between the biocatalyst and the transducer
often constitutes the major hurdle in the development of a practical device. The
attachment or the immobilization of the biocatalyst to the transducer surface is
critical to the efficiency of the biosensor. The immobilization process must produce durable and repetitive binding of the biocatalyst to the transducer without
impairing the activityor specificity of the biomolecule for the target analyte. This
biocomponent element of the biosensor may be protected by a membrane that is
permeable to certain molecules in the environment. In many cases, the biological
receptor is directly immobilized or adsorbed onto the surface of the transducer.
a generalized biosensor is shown
in Fig. 1. A
A schematic drawing for
target analyte (illustrated by solid circles) in the external medium (the sample)
must be able to enter the biosensor. The external membrane of the biosensor
must be permeable to the analyte and,if possible, exclude other chemical species
that the biosensor might also be sensitive to. The biological element inside the
biosensor then interacts with the analyte and responds in some manner that can
be detected by a transducer. The biological element may (a) convert the analyte
to another chemical species (represented by open circles) through a biochemical
reaction, (b) produce or release another chemical product in response to the analyte stimulus, (c) change its optical, electrical, or mechanical properties,
or (d)
make some other response that can be reliably quantified. There may be another
internal membrane near the transducer, which might have different permeability
a biosensor deproperties than the external membrane. The output signal from
pends on the type of transducer it uses. The transducer may be a conventional
electrochemical sensor or may be based on another technology.
It is quite evident that advances in both biotechnology and electronics have
accelerated the development of biosensors. The biotechnological advances include an increased understanding of biomolecules and biomolecular interactions.
340
Deshpande
Blologlcal
element
Semipermeable
membrane
medium
Electrolyte
Signal
/
Transducer
Internal
membrane
Fig. 1 General principlesof a biosensor. A specific chemicaltarget (analyte, represented

by solid circles)is recognized by the biological element,creating a stimulus (opencircles)
to the detecting transducer, from which a reproducible signal is measured.
The availability of morerefined and improved transducer elements is largely due

to the progress made in the microelectronics, microfabrication, and communication industries. The biosensor technology thus encompasses the molecular recognition and amplification inherent in many biochemical reactions and combines
them with the signal processing and transmission capabilities
of electronics and
fiberoptics technology. By providing valuable real-time information, biosensor
technologies have the potential to revolutionize a wide spectrum
of biochemical
analyses for health care, food-processing, agricultural, and pollution-monitoring
applications.
The developmentof a biosensor, therefore, requires an integrated, multidisciplinary team of biologists, chemists, physicists, engineers, and computer experts. This blend of expertise is generally not found in every establishment, so
biosensor developments have primarily resulted from interinstitutional collaborative projects.
111.
CHARACTERISTICS OF AN IDEAL BIOSENSOR
A principal impetusin progressing towards the development of biosensor systems

has been the need to produce a simple,
very rapid, sensitive, and easy-to-use
Biosensors
341
analytical system that doesnot need trained specialists to produce reliable results.
The aim isto develop a small, portable unit into which the sample canbe inserted
without preprocessing. The result should be obtained within seconds,
and the
answer should not be subject to interference or modification by the matrix, the
user, or environmental factors.
The optimum design of a biosensor is dictated by several basic physical
properties of the measuring system, as well as those of the media in which the
measurement is made (10,l I). Some of the most pertinent properties and characteristic behaviors of an ideal biosensor are described below.
A.
Sensitivity
The sensitivity is usually defined

as the final steady-state changein the magnitude
in concentration of a
of the biosensor output signal with respect to the change
specific chemical species (10). In other instances where measurements are based
on the dynamic response of the biosensor, it is the change in the signal with time
for a given change in concentration or some other relationship that depends on
time. Time integration, frequency analysis, or other data processing of the time
varying signals may also be of value in relating them to the concentration of the
analyte.
Factors that affect the sensitivityof a biosensor forits target analyte include
the physical size of the sensor, the thickness of the membranes and resulting
and
mass transport of chemical species from the sample to the sensing region,
various processes that deactivate the biosensor or otherwise impairits operation
over time. Ideally, the sensitivity of a given biosensor should remain constant
during its lifetime and should be sufficiently high to allow convenient measurement of the transducer output signal with electronic instrumentation.
B. Calibration
An ideal biosensor should be easily calibrated
by exposing it to prepared standard
solutions or gases containing different known concentrations
of the target analyte.
It is preferable to perform a calibration procedure only once to determine the
sensitivity of the biosensor for subsequent measurements. However, in practice
it is usually necessary to make periodic calibrations at regular intervals to characterize changes in the sensitivity with time.
Linearity
C.
A biosensor need not be linear to be practically useful, as long as the calibration
curve can be obtained with sufficient accuracy to interpret the biosensor signal.
Ideally, the measurements should be made in the linear range of the calibration
curve.
342
Deshpande
D.DetectionLimit
Ideally, the range of expected analyte concentrations for an intended application
be useful. Theoretically, the
mustbe detectable in order for the biosensor to
lowest limit of detection is dependent upon the resolution of the electronic instrumentation used for the measurement.In practice, other considerationsmay often
cause the lower limit of detection to be higher. For example, electrochemical
transducers using potentiometric measurements
may have interference from other
ions and surface reactions that limit the measurement.
E. BackgroundSignal
The background signal may make the determination of the lower limit of detection for a biosensor difficult to determine accurately. Current leakage, small potential differences in the electronic instrumentation or due to dissimilar metalto-metal contacts in the wire leads of the biosensor, or electrochemical factors
are often the major causes for the background signal.
F. Hysteresis
If the outputof a sensor follows different paths during increasing and subsequent
decreasing inputs,it is said to suffer from hysteresis.In essence, it can be considered as an effect of measurement history. An ideal biosensor should not be affected by its past history of measurements and would have zero hysteresis. Hysteresis could also affect the transient responses of the biosensor. In some cases
it can be minimized by making slow changes.
G. DriftandLong-TermStability
Drift is a slow change in output without any change
in input quantity, and it
directly relates to the long-term stability
of the measurements takenwith a biosensor. An ideal biosensor should have zero
driftand constant sensitivity for its
It
entire lifetime, or at least during the time the measurements are being made.
is usually necessary to recalibrate the biosensor at frequent intervals during its
for thedrift in sensitivity with time.
use, so thatthesignalcanbecorrected
Generally, drift is approximately linear with time between calibrations (10).
H. Specificity(Interference)
The response of a biosensor should be specific to only the changes in concentration of the target analyte andnot be influenced by the presence of other chemical
Biosensors
343
species. But normally, variations in other analytes and/or environmental factors

will affect the output. Such effects (considered as noise) should be minimized.
1.
DynamicResponse
The physical properties and relative size of a biosensor determine how quickly
it will respond to a change in concentration of the target analyte it measures
(10,ll). The principal mechanism that affects the dynamic response
is usually
simple diffusion of the chemical species from the sample to the active surface
of the transducer. Itis desirable that the response be as fast as possible (but
comparable to the speed of data acquisition).
J.
Flow Sensitivity
Generally, the response time of a biosensor decreases as flow and convective

mass transport increase. A biosensor that is not operating in a diffusion-limited
mode may be subject to measurement errors (stirring artifact) when flow is increased.
K. TemperatureDependence
Since all the physical properties are dependentupon temperature, biosensor measurements should ideally be made under isothermal conditions. Special efforts
should be made to minimize effect of varying temperatures on the output.
L.Signal-to-NoiseRatio
A large signal-to-noise ratio is an indication that the biosensor is responding
strongly to the analyteof interest, and the other extraneous effects are
kept rather
small. The signal-to-noise ratioof the biosensor measurements can be improved
by using digitalfiltering techniques to clean up the signal,
by ensemble averaging
that is synchronizedwith the repetitive stimulus,by reducing the cutoff frequency
of the amplifier, or by passing the output through analog filters.
M. Lifetime
The biological elementsused in a biosensor are generally the least stable components of the system. The lifetime of a biosensor may be dependent on the total
number of measurements made or may depend on the magnitude of the analyte
concentrations measured. Higher concentrations may lead to more rapid losses
in sensitivity. It may also be necessary to store the biosensor under refrigeration,
Deshpande
344
or the biological element may need to be supplied with a specific chemical environment to maintain its bioactive properties.
also need to be
In addition to the above properties, several other factors
considered during the design and development of biosensor products, including
the following:
Instruments should be small, self-contained, cheap and robust, capable of
interfacing with existing central laboratory systems.
The user interface should be simple for use by even unskilled operators.
No volumetric measurement of the specimen, e.g., by pipette, should be
necessary.
The test specimen should be the only addition; no further reagent addition
should be required.
Results should be unaffected by the test specimen matrix.
The time between presentation of the specimen and final result should be
less than 5 minutes.
Built-in standardization and controls are required.
Results obtained need to be quantitative, with a printed copy of the results
available.
The detection limit should be subnanomolar.
A wide analytical range is required, with
a capability for immunochemistry,
clinical chemistry, enzymology, DNA probe measurements, and a variety of other applications.
The potential for simultaneous measurement of multiple analytes should
be considered.
There should be a good correlation of results with existing test methods.
Biosensor consumables must be cheap to manufacture in bulk.
IV.
BIOLOGICALTRANSDUCERS
As described earlier, the biorecognition element
is the definitive component of

the biosensor. It is also the most crucial, being responsible for the selective recognition of the analyte, generating the physicochemical signal monitored
by the
physical transducer and, ultimately, the sensitivity of the device (12). The bioreas typified
cognition element can be categorized into two distinct types: catalytic,
by enzymes, and irreversible or affinitive, of which antibodies and the receptors
a third group, amplified or hybrid
are the best known examples. Occasionally,
configurations of the catalytic and affinitive types, can also be distinguished.
The catalytic types of biorecognition systems include single-enzyme systems, multiple enzymes, organelles, whole cells or organisms, and slices of animal or plant tissue, the latter typically containing numerous enzymes and various
Biosensors
345
cofactors with which they function more efficiently. Enzymes

still remain the
most popular choice, providing a high level
of amplification of the biorecognition
process with a high degree of selectivity. The most important enzymes from an
analytical viewpoint are the oxidoreductases, which use oxygenor NAD to catalyze the oxidation of compounds, and hydrolases, which catalyze the hydrolysis
of compounds (13,14). For example, glucose oxidase catalyzes the oxidation of
glucose to gluconic acid, which is the basis of glucose monitoring for diabetics.
Singleenzymesystems,however,
maynot alwaysproveideal,through
either instability via thermal denaturationor interaction with products or intermediates. They can also lose their activityin the immobilization step. To overcome
these limitations, multienzyme-detection systems often can be used to regenerate
an expensive coenzyme or to produce a detectable product not available from a
single enzyme system. The use of plant or animal tissue slices or whole cells
can often permit a complex sequence of reactions because coenzymes and other
cofactors are present in a more stable and natural environment. Although a loss
of specificity is often seen in such systems, the detectionof a species rather than
a specific analyte may be advantageous in some applications.
Unlike the catalytic receptors, the affinity class of biorecognition element
is even more specific in nature of the binding. These biocomponents are often
more suited to a single detection use rather than monitoring applications, since
the binding is essentially irreversible. lmmunoreceptors (antibody-antigen) are
the dominant type of affinity receptors as far as biosensor applications are concerned. Others include G-protein coupled receptors and lectins, which also exhibit
a high degree of specificity and affinity.
Biological transducers can be immobilized on a solid support in a variety
of ways. There are a number of requirements that the immobilization technique
must satisfy if biosensors are to be of practical use. These include the following:
The biological component must retain substantial biological activity when
attached to the sensor surface.
The biological film must remain tightly associated with the sensor surface
while retaining its structure and function.
The immobilized biological film must have long-term stability and durability.
The biological material needs to have a high degree of specificity to a particular biological component.
Physical adsorption, entrapment in a gel or polymer, covalent binding to
a carrier, and cross-linkingof proteins are techniques routinely usedin biosensor
development (7). The immobilization matrix may function purely as a support
or may also be concerned with mediation of the signal transduction mechanism.
Physical adsorption of the biocomponent based on van der Waals attractive forces
is the oldest and simplest immobilization method. A major disadvantage of this
Deshpande
346
method, however, is that the binding forces between the biocomponent and the
in matrices
support cannot be controlled easily. Entrapping the biocomponent
such as gels, polymers, pastes, or inks considerably improves its stability. However, to improve the performance of the biosensor, it may be essential to covalently link the biological transducer to the solid support. A prerequisite for these
immobilization procedures is the availability of functional groups on the protein
as well as on the solid support. Covalent binding
of the biocomponent to the
solid phase often results in greater precision and accuracy of the assay by eliminating within- and between-assay variation.
V.
PHYSICALTRANSDUCERS
Almost all biosensor technologies fall into four methods of signal transduction:
optical, mass, electrical, and thermal. The nature of the sensor surface, the sensors ability to measure in real time, and whether the sensor can be reused
tend to be distinctive in each method. Similarly, the merits of each method are
dictated by the application and breadth that it brings to the diagnostics industry
compared to technologies and products currently in the marketplace. The basic
characteristics of these four classesof signal transduction technologies are briefly
described below.
A. Optical Transducers
Signal transduction by monitoring changes in optical properties can be measured
by fluorescence, absorbance, refractive index, interferometry, diffraction,or polarization. Optical transducers can
be broadly divided into two further categories:
extrinsic and intrinsic. In the extrinsic mode, the incident light passes through
the sample phase and interacts directly with the sample. In the intrinsic mode,
the incident wave is not directed through the bulk sample but propagates along
a wave guide. It thus interacts with the sample only at the surface within the
evanescent field.
6. MassTransducers
The surface acoustic and piezoelectric methods of mass signal transduction are
primarily based on the measurement of interdigitating arrays of the planar devices. This type of biosensor is sensitive to changes in density, elasticity, or
electrical conductivity of the surface on whichthe acoustic wave propagates.
Piezoelectricsystemsareessentiallymicrobalancesbased
on quartzcrystals.
These crystals are mechanically distortedwhen subjected to an electric potential.
347
Biosensors
C.ElectricalTransducers
There are six typesof electrical transduction: potentiometric, amperometric, conductimetric, capacitance, impedance, and oxidation state. Of these methods, potentiometric (monitoring pH changesof the medium) and amperometric (monitoring the redox potential) have found the widest applications in the biosensor field.
D. ThermalTransducers
Since most enzyme- or microorganism-catalyzed reactions are accompanied by
considerableheatevolution,thermaltransducers,whichmonitortemperature
changes by thermistor devices, may find broader applications than other physical
transducers in the food industry.
A summary of various transducers used
in biosensor construction, with
their potential advantages and disadvantages, is presented in Table 3.
Table 3 AComparisonofTransducersUsedinBiosensorConstruction
Transducer
Sluggish response, requires a
stable reference electrode.
susceptible to electronic
noise
Simple, high selectivity
Low sensitivity
Simple, high selectivity
Low sensitivity
Remote sensing. low cost.
Interference from ambient
miniaturized, can be free
light, requires high-energy
from electrical interference
sources, only applicableto a
narrow concentration range
Fast response, simple stable
Low sensitivity in liquid apoutput signal, low cost for
plications, interference rereadout dcvicc. 110 spectal
sulting from nonspecific
sample handling
binding
Vcrsatility, free from optical
Expensive. cumbersome. reinterferences such a s color
quires a large amount of
and turbidity
enzyme
Low cost. mass production, sta-Temperature-sensitive, fabrible output. rcquires very
cation of different layers
small amount of biological
on the gate is not pcrfccted
material. can monitor several analytes simultaneously
Ion-selective electrodes Simple, reliable
Oxygen electrode
H 2 0 zelectrode
Optical systems
Piezoelectric
Calorlmetric
Ion-selective tieldeffect transistor (ISFET)
Deshpande
340
VI.
BIOSENSOR TYPES
Biosensors can be classified based on either the biocomponent used (e.g., enzymatic, immunological, or cellular) or the signal transducing system (e.g., potentiometric, amperometric, optical, thermal, etc.). The latter classification is used
in this discussion.
A.
Optical
Optical biosensors make use of the physical properties of light in various ways
to detect small changes in analyte concentration. Based on the principle used,
as surface plasmon resooptical biosensor devices could be broadly classified
nance (SPR), total internal reflection (TIR), or photon correlation spectroscopy
(PCS) instruments.
SPR instruments containing an antibody layer detect minute changesin the
refractive index on binding of the antigen (15). These changes are related to the
size of the antigen as well as the conformational change of the antibody-antigen
complex on binding. The latter involves solvent reorganization and protein unas a shift in the angle
folding. The change in the refractive index is then detected
of total absorption of incident light on a metal layer (usually gold or silver) carrying the antibodies. This system can be usedas a biosensor because the surface
plasmon resonance, which extends horizontally along the metal surface for about
100 nm, has an associated, exponentially decreasing, evanescent
field that extends
a few hundred nanometers. When
vertically into the surrounding medium for
molecules bind at the surface within the evanescent field (antibodies are about
10 nm in diameter), they perturb the field and hence the surface plasmon resonance, which therefore alters the angle of the light at which SPR occurs.
A schematic diagram of an SPR device consisting of a prism on a glass
slide carrying a thin metal layer is shown in Fig. 2. SPR-based instruments have
been successfully used for the characterization of dye monolayers (1 6), for the
measurement of antibody concentration in liquids (15), and for the binding of
the anti-human serum albumin without prior incubation or separation steps (17).
Inexpensive immunosensing devices based on SPR technology could be manufaca metalized diffraction grating on
a
tured using holographic techniques using
plastic support (18). These devices could be used for the detection of microbial
pathogens, toxins, and pesticide residues in plant and animal food products.
TIR-based biosensors detect internal reflections in a light guide. The guide
surface is coated with antibodies that come in contact with the analyte (7,s).TIR
immunosensors make use of the evanescent wave penetrating only a fraction of
a wavelength into the optically rarer medium when light coming from
an adjacent
denser medium is incident on the interface with an angle above the critical angle
(Fig. 3). Changes in the surface refractive index or absorptivity then reduce the
349
Biosensors
Reflected
light
Incident
light
- Glass substrate
- Metal
Sensitizing layer
( antibody )
""""""""
""""""""
""""""""
""""""""
""""""""
Sample solution
""""""""
""""""""
""""""""
( antigen )
""""""""
""""""""
""""""""
""""""""
Fig. 2 Biosensorsbasedonsurfaceplasmonresonance
(SPR) principle to measure
analyte/antigen concentrations in solution kinetically without prior incubation or separation steps. The device measures changes in the refractive index upon analyte-antibody
binding.
transmission of light through the guide. Kress-Rogersand Turner (19) suggested

the use of fluorescent techniques with TIR immunosensors, since the fluorescent
at the surfaceis coupled
evanescent wave originating from fluorescent complexes
back into the guide. This gives a high fluorescence intensity at the angle of total
internal reflection.
"_""""""_
_""""""""
""""""""
""""""""
""""""""
""""""""
""""""""
Light
~-guide
Incident
light
Sample solution
Sensitizing layer
-Filter
Detected
light
Fig. 3 A schematic diagram ofa biosensor based on total internal reflection (TIR) geometry. Based on the evanescent wave principle, it measures the effects of analyte binding
to its antibody or receptor immobilized on the surface.
350
Deshpande
A TIR immunosensor based on fluorescence evanescent wave technique

(fluorescence capillary-fill device, or FCFD) is commercially being marketed by
Serono Diagnostics Limited (Allentown, PA) (20). This FCFD system is capable
of achieving nanomolar detection limits and can be used for immunological detection of both small and large analyte molecules. Disposable optical devices based
on this principle, with automatic definition of sample volume by capillary tubes,
have been patented by Hirschfield (21).
PCS optical techniques for particle sizing in the submicron range measure
the time constantof the Brownian motion of particles by dynamiclight scattering
(7). In the diagnostics industry, PCS devices could be used
to monitor the changes
in particle size as a result of agglomeration following the formation of antibodyantigen complex.
B. Acoustic
Acoustic biosensors sensitive to changes in the density, elasticity, or electrical
conductivity of a surface on which the surface acoustic wave (SAW) propagates
have potential in agriculture and veterinary diagnostics for monitoring food quality. The variety and ingenuity of these
devices is extensive. Two extensive reviews cover the use of piezoelectric transducers in detection of mass (22). The
methods rely, in general, on measuring the changes in vibrational resonant frein mass on
quency of piezoelectric quartz oscillators that result from changes
oscillators surface.
A SAW immunosensor consists of a piezoelectric crystal such as quartz or
lithium niobate carrying thin-film interdigital electrode arrays. Radiofrequency
excitation of the electrode pair creates a synchronous mechanical surface wave
that is propagated on the surface of the piezoelectric substrate and recorded by
either another electrodepair on a SAW delay line or
by the same pair after refleca chemical
tion on a SAW resonator device (7,19). A schematic diagram for
sensor based on a SAW delay line is shown i n Fig. 4.
Other acoustic devices such
as a quartz microbalance or a piezoelectric
oscillator, based on the dependence
of the resonance frequency of a vibrating
quartz crystal on its mass, can also be used as biosensors. The quartz crystals
in a liquid, and if the
oscillate when immersed either partially or completely
solution properties such as density, viscosity, and conductivity are kept constant,
the change in mass of the quartz crystal can be detected (8,23). Imrnunosensors
based on the piezocrystal microbalance principles have already been commercially introduced (24).
Using conventional antibody-antigen interactions on thin substrates, detection limits of nanomolar down to picomolar concentrations have been claimed
with ideal test samples, although this is perhaps unlikely to be achieved with
normal food samples. Manufacture of the devices to produce cheap, disposable
351
Biosensors
Sample channel with
sensitizing layer for
specific bindingof
the analyte
lnterdigital
electrodes
- I' 1
Piezoelectric
substrate
SAW
delay
Fig. 4 Chemical sensor based on a surface acoustic wave (SAW) delay line.
units with uniform characteristics is also likely to pose problems, and provision
of multianalyte analysis on a single unit will be a considerable challenge.
C. Semiconductor
Field-effect transistor (FET) devices similar
to the ion-sensitive FET devices(ISFET) have been used for the detection of minute potential changes associated
with the formation of an antibody-antigen complex (25). Such direct immunosensors are called IMFETs. ISFET and other microelectronic gas or ion sensors
can also be used in indirect electrochemical immunoassays using enzyme labels
instead of the traditional ion-selective electrodes of gas probes (8,19,26). The
fast response,good signal-to-noise ratio,and small sizeof FET devices are particularly useful in flow-injection analysis techniques. Because of manual handling
in conventional ELISAs (chloramphenicol acetyl transferase, or CAT assays),
the variations in the binding time between the reactants need to be eliminated
by extending the timeso that slight variationswill not influence thefinal outcome.
In contrast, in flow-injection analysis, the only factor
that varies is the time during
Deshpande
352
which the pulse of the substrate is introduced. This makesit possible to keep the
variation between assays to less than 2%. Potentially, the FET devices can be
constructed as inexpensive disposable units, although their full potential has not
yet been explored by the food diagnostics industry.
A schematic diagram of a prototype IMFET devicein which the metal gate
is replaced by an antibody or antigen immobilized on a membrane is shown in
is controlled
Fig. 5. The conductivity of the n-channel region in the p-type silicon
by the strength of the electric field at the gate and is measured by the application
its proper
of a voltage between the source and drain electrodes. A prerequisite for
functioning, however, is to have the solution-membrane interface remain ideally
polarized and thus impermeable to the passageof the charge. Failure to meet this
criterion results in the poor sensitivity of the instrument.
Yet another semiconductor device, TELISA (or thermistor ELISA), has
great promise in the diagnostics industry. The heat-sensitive enzyme thermistor
in a small,wellmeasurestheheatevolved
in anenzyme-catalyzedreaction
of an enzymethermistor is shown
insulatedchamber.Aschematicdiagram
in Fig. 6. A small sample volume is injected
in a comparatively rapid buffer
stream. A short sample pulse results in a temperature peak that is proportional
a linear operating
to the concentration of the analyte (enzyme substrate) over
range of concentration. Normally, the thermograms are evaluated by peak height
Encapsulant
Membrane
insulator
"""""
n - Si source
p-Type
silicon
substrate
n Si drain
<
Antigen
Immobilized
antibody
Fig. 5 Immuno-Field-Effect Transistor (IMFET) sensor for direct potentiometric monitoring of the antibody-antigen complex.
353
Biosensors
Polyurethane
//
Heat Exchanger
\
Thermostatted
Insulation
Aluminum Block
Auxiliary or
Column
Reference
Column
Enzyme
Fig. 6 A schematic representation of an enzyme thermistor that usesflow injection analysis for analyte detection.
determination. Continuous sample introduction leadsto a shift of the temperature

this shiftisalsoproportionaltothe
signal to a newsteady-statelevel.Since
analyte concentration, the instrument can easily be used for continuous monitoring.
The system shown in Fig. 6 contains two identical flow lines, which can
be independently used for two different enzyme columns for different analyses.
Alternatively, one of the columns can be used as a reference column to compensate for any nonspecific reactions. In normal operation, however, the temperature
of the enzyme columnis registered versus the temperature measured witha reference thermistor mounted in the heat sink of the calorimeter.
as albumin, gentaTELISA hasbeen used for the detection of such antigens
micin, insulin, insecticides, and heavy metal ions, and the procedure has been
automated for monitoring hormones and proteins produced by biotechnological
methods (27,28).
D. Electrochemical
Electrochemicalbiosensorssimilartoenzymeelectrodescan
also beusedin
developing rapid assays for food quality control. They have a unique blend of
sensitivity with simplicity that resultsin a family of low-cost, rapid, and portable
chemical sensors capable of operating in turbid solutions such as food analytes
(29).
Deshpande
354
Electrochemical transducers can be divided into two general categories:

potentiometric devices that require the derived voltage to be determined with
reference to a second electrode under conditions of essentially zero current flow,
and amperometric devices that measure the current flowing between two electrodes in response to the application of a defined voltage. Potentiometric devices
measure a logarithmic response and are critically dependent on a precise reference electrode. The responses, however, are largely independent
of mass transport
effects. In contrast, amperometric electrodes produce a linear response and are
relatively insensitive to fluctuations in the reference voltage (7,8,19). However,
since the species being measured is consumed, they are affected by changes
in
the rate of diffusion of the analyte to the electrode.
Janata (30) first reported that the changes in electrical charge on antibodyantigen interaction could be detected directly using a simple potentiometric system. Subsequently, Yamamoto et al. (3 1 ) covalently immobilized anti-human
chorionic gonadotrophin (anti-hCG) on a cyanogen bromide-activated titanium
a titanium reference electrode
wire and monitored the potential difference against
on addition of hCG (Fig.7). These researchers suggested thatit would be feasible
e5mV
I
"
Reference
Electrode
Titanium Wire
Electrode
Human chorionic gonadotrophin (hCG)
Immobilized anti-hCG
Fig. 7 A direct potentiometric immunosensor based on a titanium wire electrode.
Biosensors
355
to monitor voltage differences as low as 0.05 mV using their apparatus, although

they did encounter problems resulting from nonspecific binding. IMFET devices,
described earlier, also fall in the category of direct potentiometry-based immunobiosensors.
7.
Potentiometric
Potentiometric membrane electrodes, such as pH, solid-state, polymer, and gassensing electrodes, have been used as internal sensing elements for biosensors.
A typical classification of these ion-selective electrodes
(ISE) is shown in Table 4.
Gas-sensing electrodes are by far the most popular choice because of the inherent
selectivity of gas-permeable membranes (32,33).
Potentiometric biosensors based on gas-sensing membrane electrode consist of a hydrophobic gas-permeable membrane anda pH electrode separatedby a
thin layer of internal electrolyte. The differences between the several gas-sensing
electrodes are mainly apparent with respect to their internal filling solutions and
their operating pH ranges. Gas-sensing electrodes are quite selective since their
membranes are permeable only to gases.Samplesolutionshavetobetotally
aqueous since the hydrophobicity of these membranes would be compromised
by organic solvents, thereby leading to membrane failure.
The usefulness of gas-sensing electrodes is often hampered by pH limitations imposed by samples or by the biocatalysts used in the construction of biosensors. Compromises often have to be made in order to satisfy both the pH
requirements of the electrodes and those of the biological components, with the
latter usually having stringent pH limitations due to their biological nature. The
pH requirements of the gas-sensing electrodes themselves arise from the fact that
the electrodes are responsive onlyto the gaseous formof the measured substrate.
Ion-selective electrodes, however, can tolerate a certain degree
of turbidity in
samples and are thus especially suited for food analysis.
Someexamples of contemporarypotentiometricmembraneelectrodes
are summarized in Table 5. The lifetimes of these biosensors vary from hours
Table 4 Classification of Ion-Selective Electrodes Based on Sensor Membrane

Composition
Type of electrode
Composition
Glass
Solid state
Liquid ion-exchange membrane
With coating over membranes of
ion-selective electrodes
ofmembrane
sensor
H', Na', monovalent cations

F-, Cl-, Br-, I-, CN-, SCN-, Cd?+, Cd', Pb2+
Ca", CI-, divalent cations, BFJ-, N03-, cIoJ-,K+
C02, NH3, H2S, NO2, SOz
Table 5 SelectedExamplesofPotentiometricBiocatalyticMembraneElectrodes
Internal
sensing
Lifetime
Substrate
Ref.
Enzyme electrodes
Adenosine
L-Alanine
Glutamine
Guanine
Gluconate
Histidine
Salicylate
Tyrosine
Bacterial electrodes
L-Arginine
L-Aspartate
L-Cysteine
L-Glutamate
Glutamine
L-Histidine
NAD+
Nitrate
Nitrilotriacetic acid
Pyruvate
Serine
Sulfate
Sugars
Tyrosine
Uric acid
Plant tissue electrodes
Ascorbate
Cysteine
Dopamine
Glutamate
Phosphatelfluoride
Pyruvate
Urea
L-GlutaminehAsparagine
Tyrosine
Spermidine
Animal tissue electrodes
AMP
Guanine
Glucosamine 6-phosphate
kidneyPorcine
Glutamine
(d)
element
Biocatalyst
Adenosine deaminase
L-Alanine dehydrogenase
Glutaminase
Guanase
Gluconate kinase-6-phospho-~gluconate dehydrogenase
Histidine decarboxylase
Salicylate dehydroxylase
L-Tyrosine decarboxylase
Streptococcus fueciutn
Bacterirtm caduveris
Proteus morgnnii
Escherichia coli
Sarcina jlava
Pseudomonas spp.
Esclzerichia colilNADase
Azotobacter vinelundii
Pseudomonas spp.
Streptococcus faecium
Clostridium ncidiurici
Desulfovibrio desulfuricans
20
Bacteria from dental plaque

Aeromonas phenologenes
Pichi tnernhranae fnciens
IO
1
42
4
30
12
0.42
IO
6
21
14
21 s
7
14
30
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
14
51
52
IO
53
54
3
8
55
50
56
Cucumber
Cucumber leaf
Banana pulp
Yellow squash
Potato tuberlGlucose oxidase
Corn kernel
Jack bean meal
Magnolia flower
5-6
57
28
58
59
7
28
7
94
14
60
Sugar beet
Pea seedling
8
20
65
66
Rabbit muscle
Rabbit liver
Porcine kidney
28
14
21
67
68
69
30
70
10
61
62
63
64
Biosensors
357
to weeks, depending upon the biocomponent used in the preparation of the electrode.
2. Amperometric
Amperometric biosensors are usually based on oxidoreductase enzymes, which
commonly use oxygen, NAD', or NADP' as the electron acceptor that recycles
as the electron
the enzyme after the substrate reaction. Enzymes using oxygen
acceptor are commonly called oxidases, while those usingNAD' or NADP' are
called dehydrogenases or reductases. Of these two groups, oxidase enzymes are
the most commonly used as biocomponents.
The first amperometric biosensors used the Clark oxygen electrode as the
physical transducer. The oxygen electrode is generally constructed using a twoelectrode system, e.g.,a platinum cathode and a silver anode. A gas-only permeable membrane (e.g., Teflon) covers the tip of the cathode. A biochemical layer
(most often an enzyme) is then placed on the oxygen electrode and secured with
a coarse membrane that allows the passageof analyte, thus forming a biosensor.
A negative potential is then applied to reduce oxygen at the cathode. The reduction current is directly proportional to the oxygen concentration. In the presence
of substrate, the enzymatic reaction decreases the oxygen concentration, and the
decrease in the oxygen reduction current is proportional to substrate concentration.
Theprincipaladvantage of anoxygenelectrode-basedbiosensor
is its
excellentselectivity. The selectivityoftheenzyme
is not compromised by
the oxygen electrodes, because only gases can pass through the membrane and
only substances that are reducible at the applied potential can interfere.
The limitations of this type of biosensor are that it is sensitive to ambient oxygen concentrationsandthegas-onlypermeablemembranecanbecomeblockedor
clogged, which often necessitates the use of a second coarse membrane (e.g., a
dialysis membrane). As the membrane layers become thicker and more complex,
the response and recovery times increase, and the biosensor may require more
frequent calibration (71). The maintenance of calibration for many biosensors is
dependent not only on enzymatic activity but also on the constancy of the diffusional pathways for reactants and products. The thicker and more complicated
the biochemical layer and its associated membranes, the more difficult
it is to
maintain constant diffusion of reactants and products, thus increasing the frequency of calibration.
Oxygen electrode-based biosensors have been used to determine a variety
of enzymatic substrates including glucose, monosaccharide, hypoxanthine,
lac(8).
tate, amino acids, sulfite, salicylate, oxalate, and pyruvate
Hydrogen peroxide-and dehydrogenase-based biosensors also operate on
similar principles.The former functionby oxidizing hydrogen peroxide,a product
358
Deshpande
of the enzymatic reaction, while the latter utilize the enzymatic oxidation of
NADH or NADPH.
The mainproblemwiththedehydrogenase-basedbiosensors
is thatthe
NADt/NADH and NADP+/NADPH redox couples are not always completely
reversible, and when the reduced formis reoxidized, some product can form. this
oftenirreversiblyadsorbsontheelectrodesurface,causingelectrodefouling.
After repeated use, the oxidized cofactor decreases
in concentration if the electrochemical recycling is not 100% efficient or if it is not confined to the electrode
surface. Thus, it may become a limiting factor in the enzymatic reaction. In conjunction with this problem, if the undesired product fouls the electrode surface,
the current response of the electrode decreases with time because the active electrode area is decreasing. Careful attention to the choice of electrode material and
electrolysis conditions can minimize these problems.
Alternatively, a mediator with better electrochemical properties can be used
to recycle the NAD'/NADH redox couple. This approach was used
in the construction of ethanolandlacticacidsensors,
in whichflavinmononucleotide
(FMN) was the mediator, and its reduced form, FMNH?, was oxidized at a platinum electrode, producing a current directly proportional to substrate concentration (72).
Amperometric biosensors that use oxidase enzymes normally rely on oxygen as an electron acceptorto recycle the enzyme after conversion of the substrate
to the product. The upper linear rangeof these biosensors is limited at high substrate concentrations because oxygen becomes a limiting factor due to its limited
solubility. As the substrate concentration increases, more oxygen is used
in the
enzymatic reaction; when the concentration of the oxygen becomes too low, it
becomes a limiting factor. Consequently, each increasein substrate concentration
results in less of an increase in the overall reaction rate, and the biosensor response begins to level off, eventually becoming independentof substrate concentration and limited by the oxygen concentration. Also, hydrogen peroxide proin sufficient concentration, is known to
duced from the oxygen, when present
deactivate many enzymes (7 1).
Electron mediators can be used
to alleviate both problems by replacing
oxygen as the electron acceptor. This approach usually extends the linear range
and often lowers the working potential, thereby reducing noise, background curto interferences.Byeliminatinghydrogenperoxide,the
rent,andsignalsdue
operating lifetime of the biosensors is often extended. The commonly used electronmediatorsincludetetrahiafulvalene,hexacyanoferrate,N-methylphenazinium, and ferrocene and its derivatives (7). Ferrocenes are popular because they
are hydrophobic, exhibit good electrochemistry, and can be structurally altered
to change their redox potential. Hydrophobicity is important because it prevents
the mediator from leaching from the electrode surface when used
in aqueous
systems.
Biosensors
359
Table 6 ApplicationsofBiosensorsintheFoodIndustry
Compound
measured
Applications
Amino acids
Alanine, arginine, asparagine, aspartic acid, cysteine, glutamine, glutamic acid, glutathione, histidine, leucine, lysine,
methionine, N-acetylmethionine, phenylalanine, sarcosine,
serine, tyrosine, tryptophan, valine
Amines,amides,hetero-Aminopyrine,aniline,aromaticamines,acetylcholine,chocyclic
compounds
line,
phosphatidylcholine,
creatine,
guanidine,
guanosine,
penicillin, spermine, creatinine, uric acid, urea, xanthine,
hypoxanthine
Carbohydrates
Amygdalin,
galactose,
glucose,
glucose-6-phosphate,
lactose,
maltose, sucrose, starch
Carboxylic
acids
Acetic
acid,
formic
acid,
gluconic
acid,
isocitric
acid,
ascorbic acid, lactic acid, malic acid, oxalic acid, pyruvic
acid, succinic acid, nitriloacetic acid
Gases
NHI, Hz, CH,, SO?, NO
NAD(P)H,
ATP,
Cofactors
AMP,
Hz02
Inorganic
ions
Fluoride.
nitrite,
nitrate,
phosphate,
sulfate,
sulfite,
Hg",
Zn"
Complex
variables
Antibiotics,
assimilable
carbohydrates,
assimilable
substances, biological oxygen demand, freshness of meat, mutagens, vitamins
Alcohols,
phenols
Acetaldehyde,
bilirubin,
catechol,
cholesterol,
cholesterol
ester, ethanol, glycerol, glycerol esters, methanol, phenol
Selected examples of the applications of biosensors in food analysis are

summarized in Table 6.
VII.
DNAPROBETECHNOLOGY
Recombinant DNA techniques and DNA probe-based diagnostics will play

an
increasingly important rolein the determination of pathogenic microorganisms in
foods. These techniques depend on the selective cleavage of DNA by restriction
endonucleases and on the localization of the specific sequences of nucleotides
after hybridization with known DNA or RNA fragments (probes) labeled with a
radionuclide (e.g., zzP, '"I, j5S, zH)or an alternative label. The latter includes
fluorescent labels (e.g., fluorescein. rhodamines, ethidium, rare-earth chelates),
luminescent labels (luminol derivatives, acridinium esters, luciferase). enzyme
of two
markers (alkaline phosphatase, horseradish peroxidase), or a combination
360
Deshpande
or more of these label systems. Therefore, the detection systemsused in the field
of immunoassay are also applicable to DNA probing.
Nucleotide-probing techniques rely primarily on the ability of the labeled
DNA and RNA probes to detect antigen by binding to the complementary arrangements of target nucleotides. Since such binding, under appropriate assay
conditions, can be extremely strong and specific, DNA/RNA probe assays can be
extremely specific and sensitive. These characteristics are mandatory for devising
sensitive and specific diagnostic assays. Furthermore, these probes can be amplified by polymerasechainreaction(PCR)usingappropriateenzymesystems,
thereby providing the starting material in great quantity for testing.The PCR is a
three-step process that duplicates a DNA fragment bound by two oligonucleotide
primers.Eachprimer is complementary to one strand of thedouble-stranded
DNA template used. The first step of the process is initiated by denaturing or
separating the double-stranded DNA target into its single-stranded components
by heating the sample to 90-95C in the presence of a large excess of the two
primers. In the second step, the temperature of the reaction is lowered to about
10C below the melting temperature of the primers, and the primers are allowed
to anneal or hybridize to their complementary sequence on the DNA template
molecule. The third step of the reaction is carried out by raising the temperature
of the reaction to 70-73"C, the optimal temperature for extension of the primers
by a DNA polymerase. One cycle of denaturation, annealing, and primer extension results in a doubling of the DNA target sequence. By repeating the cycles
as many as 25-30 times, the DNA sequence flanking the primers can be amplified
as much as a millionfold. The PCR can also be used to assay RNA in a sample
if the RNA template is first transcribed into DNA by reverse transcriptase.
Blackburn et al. (73) described an electrochemiluminescence (ECL) for the
detection of PCR
reaction
products
using
ruthenium(I1)
tris(bipyridy1)
(Ru(bpy)32+)to label the probe. ECL assays detect the hybridization of labeled
of PCR products is demonstrated
probes to nucleic acid sequences. Quantification
in Fig. 8. I n this format, the PCR is first used to amplify the specific genes by
use of two primers, one of which is biotinylated. The double-stranded DNA is
then captured on streptavidin-coated microparticles and washed with an alkaline
solution to denature and separate the strandsof DNA. Incubation of the particlebound DNA with Ru(bpy)32'-labeled probes is followed by washing the sample
and quantifying the particle-bound label.
Other recent amplification systems include the ligase chain reaction (LCR),
Q-p replicase,Syva'sPA-singleprimeramplification,andself-sustainedsequence replication (3SRJ (8).
Although the current nucleotide probe-based techniques are slow and laborious in comparison to technologies such as immunoassays, several commercial
products are already being marketed. DNA probe technology will find its general
use in medicine, clinical sciences, and food technology. In the food technology
361
Biosensors
PCR Reaction
Biotin
mmmOiigo
Polymerase
extension
Bindina to Beads
BS
Denature to Separate Strands
/"".
Hvbridize, Wash. and Read
WBb
Fig. 8 Diagram of the procedure for the detection of polymerase chain reaction (PCR)
products in the DNAprobe assay. B = Biotin, S = streptavidin-coated microparticles;
Ru(bpy),".
area, conventional microbiological testing

of a food product often requires
as
long as 4-6 days, a time frame food processors could
ill afford to wait, given
of mostfreshandmanyprocessedfoods.
DNA
thehighlyperishablenature
probes have been used for the identification
of certain bacterial cells. Other examples are the detection of Escherichia coli, Yersinia erlterocolitica, Listeria spp.,
and Salmonella in foods and drinking water.
VIII. FLOW CYTOMETRY

The flowcytometrymicrofluorimeter
(FCM), or fluorescence-activatedcell
sorter, is one of the most sophisticated machines dedicated to cell analysis and
Deshpande
362
sorting available today. The majority of applications of FCM at present relate to

biomedical research, although increasing uses are foreseenin veterinary and food
sciences, especially in microbiological quality testing.
of cell properties on
Flow cytometry simultaneously measures a number
an individual basis at speeds of up to 5000 cells per second. There is no need
first to isolate the cells of interest from the rest of the population. The cells are
usually tagged with fluorescent markers indicative of biological properties prior
to their passing individually through a focused laser beam. Strategically placed
photosensors then measure optical changes as each cell passes through the beam.
These changes are characteristic of each cell type, and the resulting signals are
analyzed by a data processing system. Cell sorting is accomplished by breaking
up the fluid stream under pressure into tiny droplets that are electrically charged
according to the type of cell that each contains. An actual physical separation is
achieved electrostatically (8).
Flow cytometry is quiteuseful in studying cell surface markers fluorescently labeled with antibodies or lectins. This allows rapid identification of subpopulations in a heterogeneous population of cells, thereby alleviating the need
for isolation of the desired cells. Any cell surface molecule can be used as a cell
population marker provided it can be made immunogenic. Twoor more antibodies can beusedtogether if they are conjugated to fluorescent dyes thathave
nonoverlapping emission spectra. A fundamental difference between FCM and
ELISA, however, is that FCM must be performed in the liquid phase, whereas
ELISA is a solid-phase assay. The former
is thus ideally suitedto microbiological
testing. FCM also allows for rapid assay turnover times of 15 minutes or less,
since the unbound fluorophore is not detected by the machine, thereby making
all the washing steps unnecessary. In addition, the binding of fluorophores used
in FCM occurs immediately, whereas the chromogenic changes
that form the
basis of ELISA take time to develop. Immunoassays using fuorophores are also
10-20 times more sensitive than colorimetric assays; thus. an approximately 50fold greater dynamic range is obtainable from FCM techniques compared to microtitration plate-based ELISA. FCM, therefore, is increasingly being reviewed
as the future technology in microbiological and clinical testing.
IX.
BIOLUMINESCENCE ASSAYS
Luminescence techniques based on the detection of light generated by enzymemediated reactions represent someof the fastest assays currently availablefor the
of the few rapid technologiesto
detection of microbial contaminants and are some
be currently utilized extensively in the food industry. Techniques based on this
phenomenon can be divided into two broad categories: ATP bioluminescence
and bacterial bioluminescence.
Biosensors
363
The ATP bioluminescence technique was first described in the 1960s by

National Aeronautics and Space Administration (NASA) scientists(74). It makes
use of the fact that all living cells contain ATP, which
is the universal energy
donor for metabolic reactions. An enzyme-substrate complex, luciferase-luciferin, present in firefly tails converts the chemical energy associated with ATP
into light by a stoichiometric reaction (Fig. 9). The amount of light emitted is
proportional to the concentration of ATP present and can be easily quantified
using a luminometer. The ATP pool within a cell is normally constant (75). A
bacterial cell typically contains one femtogram (1 X IO-" g) of ATP. Thus, the
measurement of cellular ATP is a good indicator of biomass. The light produced
during the reaction is directly related to the number of metabolically active cells
present in the assay. The sensitivity of the ATP bioluminescence assay is such
that approximately 1000 cells can be detected. Modifications
of this basic technology, using a coupled cycling reaction involving myokinase and pyruvate kinase,
can detect fewer than 100 cells (76).In another variation of this technique, cellular adenylate kinase is reacted with ADP to form ATP, which is then measured
using the firefly luciferase reaction.
Sharpe et al. (77) were the first to apply the method to the detection of
microorganisms in food, but the high levels of ATP from nonmicrobial sources
Luciferase
LUCIFERIN.AMP
PPI
Fig. 9 Basicprinciple of anATPbioluminescenceassay.ATP

= Adenosinetriphosphate; Luciferin. AMP = luciferyladenylate;Ppi = pyrophosphate;AMP = adenosine
monophosphate.
facturer
Deshpande
364
Table 7 CommerciallyAvailableATPBioluminescenceHygiene
Monitoring Kits
name Instrument/Test
Bio-Orbit"
GEM@
Hy-Lite'
InspectoPlSystem SureTM
Lightning@
Lumac@
Luminator"/PocketSwabTM
Uni-Lite@'/Uni-Lite@Xcel
Profile"
Bio-Orbit Oy, Turku, Finland

GEM Biomedical Inc., Hamden, CT
E. Merck, Darmstadt, Germany
Celsis International plc, Cambridge, U.K.
Idexx Laboratories, Westbrook, ME
Lumac bv, Landgraaf, The Netherlands
Charm Sciences, Inc., Malden, MA
Biotrace Ltd., Bridgend, U.K.
New Horizons Diagnostics, Columbia, MD
reduced the sensitivity substantially. Although there continued to be an interest

in the application of ATP bioluminescence for assessing microbial contamination
of food, it was not until the early 1990s that the technique came of age in the
food industry (78-80).
Bioluminescent methods have found several niches in the food processing
industry. Perhaps the most prominent is astool
a to monitor the efficacyof sanitation. ATP hygiene monitoring has been
used in a variety of processing situations,
including breweries, dairy plants, meat processing plants, and fruit juice opera7).
tions. Several companies now offer ATP hygiene monitoring kits (Table
Bioluminescence is also used in reporter gene technology, which is rapidly
gaining acceptance in the biological sciences as a sensitive and convenient alter(76). The
native to conventional chloramphenicol acetyl transferase (CAT) assays
bioluminescent reporter gene is linked to a geneof interest and expression of the
bioluminescent protein, e.g., firefly luciferase, marine bacterial luciferase,Rerzillu
luciferase, green fluorescent protein, etc.. used to monitor the activation of the
gene.
X.
EMERGINGTECHNOLOGIES
Overthe past two decades,enormousactivityhas

takenplace in the field of
sensor technology. Biosensors,in particular, have attracted considerable attention
because of their extraordinary sensitivities and specificities. However, such devices often lack storage and operational stability because they are based
on fragile
biological recognition elements: enzymes or antibodies. For this reason, biosenin the early euphoric
sors have not become quite the commercial success expected
Biosensors
365
development phase. Two emerging technologies, electronic nose and molecular

imprinting, however, could provide an alternative. These are briefly described
below.
A.
ElectronicNose
The world of electronic nose technology mimicking the human nose and
its utility
in the flavor industry as both a research and quality assurance tool literally exploded in the United States in late 1994. Traditionally, much of the flavor testing
in the food and beverage industries as well as flavor and fragrance manufacturers
has relied on sensory evaluation by trained expert panels, such as U.S. Food and
Drug Administration (FDA) inspectors, who assess seafood freshness, coffee and
tea tasters, or brewmasters and wine stewards. Flavor and fragrance companies
rely heavily on sensory evaluation throughout the development process to
final
production. The results areusually highly subjective, prone to error, and difficult
to relate to other analyses. Current research seems to indicate that no two humans
really perceive the same aroma quite the same, nor can they communicate a common descriptor easily due to an array of cultural, presuppositive, and associative
memory effects (81).
20-300 daltons)
Flavor molecules are generally small (molecular weight
and polar and can be detected by humans at levels below one part per billion.
To date, more than 6000 flavor-active compounds have beenidentified, and more
are being discovered each passing day. Recognizable flavors or odors arise from
the specific contribution of complex mixtures of many odorous molecules, each
of different concentration. Attempting to detect complex odors at these levels by
conventional analytical techniques is very expensive and not always possible. It
is therefore not surprising that traditional (organoleptic) methods of odor assessment have survived for so long. Analytical techniques can objectively discriminate odors, but the sample must be separated into
its individual components using
gas chromatography/mass spectroscopy ( G U M S ) techniques. While these techniques can provide a chemical tally of volatiles in a food, flavor, or fragrance,
aroma, judge its character, or
none can tell if the identified component has an
predict synergistic effects with other compounds.
In contrast, human olfaction
can discriminate aromas without separating mixtures into individual compounds.
In the human brain, a kind of pattern recognition may also take place to decode
signals sent from receptor cells located in the human olfaction system (82).
In recent years, significant interest in the use of sensor arrays to discriminate between odors has arisen. If an array of nonspecific sensors could be compiled to rival the human olfactory system, then the sample need not be separated
(83,84). Analysis would prove
and could be monitored analytically as a whole
rapid, nondestructive, and continuous. The term electronic nose (or artificial
366
Deshpande
nose) has been coined to describe such an array of chemical sensors, where
each sensor has only partial specificity to a wide range of odorant molecules,
coupled with a suitable pattern recognition system software.
In order to discriminate aromas by mimicking the human olfaction system,
attempts to construct the so-called artificial nose by utilizing various gas sensors have been made since the early
1980s (85.86). Metal oxide semiconductor
sensors (85,87-89), surface acoustic waves (SAW) sensors (90), and quartz resoits own
nator sensors (91) seem potential methods. Although every sensor has
advantages and limitations, there are
not such sensors that can show perfect selectivity to specific compounds or a group of compounds. Pattern recognition of
into every analytical
the responses from a multisensor array was incorporated
procedure due to the nonselectivity of gas sensors. Early efforts were largely
focused on detecting specific gases or hazardous compounds on the basisof their
response patterns (88-90). From the viewpoint of aroma analysis, discriminating
is essential. As the first step for
a particular gas mixture from other mixtures
developing a simple aroma-monitoring system, pattern recognition analysis for
responses to food aromas from a gas sensor array was attempted using semiconductor gas sensors (92.93). Such an array may consist of as many as six or more
semiconductor gas sensors. These are preferred over other types of sensors because of their durability, high sensitivity for most reducing compounds. and insensitivity to water vapor.
The most popular sensors used to develop electronic noses suitable for use
within the food industry are described below.
7. Semiconductor Metal Oxide Chemoresistive Sensors

Much interest has been devoted to the development of chemoresistive arrays of
inorganic semiconducting materials such as oxides and catalytic metals(94).The
oxide materials contain chemisorbed oxygen species with which interaction
of
odor molecules alters the conductivity
of the oxide. These devices operate
at
elevated temperatures (400-600C) to avoid interference from water and to aid
rapid response and recovery times. This resultsin high power consumption. They
are quite sensitive to combustible materials such as alcohols but are less good
at detecting sulfur- or nitrogen-based odors (95).
Alpha M.O.S. America, Inc. (Belle Mead,NJ) has introduced an electronic
nose based on this technology. It consists o f 6, 12, or 18 metal oxide sensors. It
is an open-ended instrument platform that has the potential to use sensors based
o n other technologies within the instrument.
2. Quartz Resonator Sensors

Electronic nose systems based on this principle consist of a piezoelectric quartz
or
crystal oscillator coated with a sensing membrane such as acetyl cellulose
Biosensors
367
lecithin. Adsorption of odor molecules on the membrane leads to changes

in
the resonant frequency of the device due to a change in mass, provided that the
viscoelasticeffectsarenegligible.Differentmembranescanbeused,andthe
changes in their resonant frequencies in the presence of different odors can be
analyzed. Electronic noses based on
this technology are commercially not yet
available.
3. Conducting Polymers
The unique electrical properties of conducting polymeric materials have been
exploited in the preparation of electronic noses. They have the advantage that a
wide variety of suitable materials exists. Two main classes are poly(pyrro1e)s
and poly(ani1ine)s. The materials are easyto process, so preparation of reproducible gas sensors is possible. One of the first applications of conducting polymers
as odor sensors was described by Persaud and Pelosi (86).
Several properties of the polymers can be used for measurement, including
changes in the work function of the polymers, as well as Inass changes that can
be monitored when polymers are depositedonto quartz resonator type substrates.
The most popular parameter to monitor, however,is the conductivity of the polymers, which is known to change rapidly and reversibly in the presence of odors.
The adsorbed odor molecules are believed to cause a swelling of the polymers
and to interfere with charge transfer within the polymer and interchain hopping
mechanisms (95). These sensors can be used at ambient temperatures and have
quite good sensitivities, typically 0.1- 100 ppm.
Electronic noses based on conducting polymer arrays are available commercially. Examples include the e-Noseo// 4000 Aroma Analysis System (NeotronicsScientific Inc., FloweryBranch, GA) andthe AromaScannero//from
Aromascan, Inc. (Hollis, NH).
The electronic nose technology does provide objective and reproducible
aroma discrimination on a wide variety of sample types with a sensitivitycomparable to human nose. Analysis times are on the order of minutes, yet the results
compare very favorably with GUMS, and sensory panel analyses that can take
much longer. There continues to be ongoing research into new sensor technologies, materials, and fabrication methods. At least one company offers multiple
hybrid systems. Toko has reported a taste sensor that has successfully discriminatedthebasic
taste qualities-sourness,saltiness,bitterness,sweetness,and
umami-in a number of food products and materials(8 I ) . The integrationof taste
sensors with electronic nose technology could truly revolutionize flavor analysis.
all
To date, all attempts to develop universal electronic nose suitable for
applications have failed. Problems associated with sampling. calibration, sensor
drift, and control of humidity and temperature are the 1na.jor causes for concern.
Improvements in associatedtechnologiessuchasnewodor-sensingmaterials,
Deshpande
368
development of hybrid noses, smarter pattern recognition techniques, and miniaturization are essential if the market potential of such an instruments is to be
realized (84,95). Ultimately, the development of application-specific noses that
are tailor-made for a given problem may prove more successful than attempts to
develop a single system suitable for universal applications.
B.
Molecular Imprinting
The technology of molecular imprinting could alleviate several problems associated with the biorecognition element of the biosensors. It provides highly stable
synthetic polymers that possess selective molecular recognition properties because of recognition sites within the polymer matrix that are complementary to
the analyte in the shape and positioning of the functional groups. Some of these
polymers have high selectivities andaffinity constants, comparable with naturally
occurring recognition systems such as monoclonal antibodies or receptors. This
makes them especially suitable as constituents in chemical (biomimetic) sensors.
Molecular recognition between a molecular receptor (host) and a substrate
(guest) in a matrix containing structurally related molecules requires discrimination and binding. This can happen only if the binding sites of the host and guest
molecules complement each otherin size, shape, and chemical functionality.
Biological systems, such as enzyme-substrate, antibody-antigen, and hormone-receptor systems, demonstrate molecular recognition properties that have developed
by natural selection.
One of the most intriguing areas for host-guest chemistry
is the development of biomimetic recognition systems. As described earlier, a wide range
of
analytical procedures depend on reliable and sensitive biological recognition elements such as antibodies, receptors, and enzymes. Because such biomolecules
can suffer from stability problems, synthetic counterparts are desirable. One such
approach to biomimetic recognition is the fabrication of molecularly imprinted
polymers (MIPS) (96).
Molecular imprinting is a powerful method for preparing synthetic recognitionsiteswithpredeterminedselectivityforvarioussubstances.
It canbeap(97)andthepreorganized
proached in twoways:theself-assemblyapproach
approach (98). These two approaches, which differ with
respect to interaction
mechanism in prepolymerization, follow common molecular recognition terminology.
The self-assembly nlolecular imprinting approach involves host-guest complexes produced from weak intermolecular interactions (e.g., ionic or hydrophobic interactions, hydrogen bonding, and metal coordinations) between the analyte
and the monomer precursors (97,99). These complexes are spontaneously established i n the liquid phase and are then sterically fixed by polymerization with a
high degree of crosslinking. After removal of the print molecules from the re-
Biosensors
369
sulting macroporous matrix, vacant recognition sites that are specific to the print
molecule are established.The shapeof the sites, maintained by the polymer backbone and the arrangement of the functional groupsin the recognition sites, results
in affinity for the analyte.
In the preorganized molecular imprinting approach, strong, reversible, covalent arrangements (e.g., boronate esters, imines, and ketals) are formed between
the monomers and the print molecules before polymerization. print
The molecule,
therefore, needs to be derivatized with the monomers prior to actual imprinting.
After cleaving the covalent bonds that hold the print molecules to the macropoto the analyte remainin the polymer
rous matrix, recognition sites complementary
(98).
MIPs have been successfully prepared with affinities for proteins, amino
acid derivatives, sugars and their derivatives, vitamins, nucleotide bases,
pesticides, and a wide varietyof pharmaceuticals (98). The binding of some of these
of some natural monoclonal
polymers has been comparable with the binding
antibodies ( 100,lO1).
The primary advantageof MIPs is that imprints can be made
of compounds
against which it is difficult to raise high-affinity antibodies. Other advantages
include their long-term stability and resistance to chemically harsh environments
( 102).
MIPs have unique properties that make them especially suitable for sensor
of medical,
technology. They exhibit good specificity for various compounds
environmental, and industrial interest, andthey have excellent operational stability. Their recognition properties are unaffected by acid, base, heat, or organic
in
phase treatment (102), making them highly suitable as recognition elements
chemical sensors. Some examples are shown in Table 8. To date, the most convincing demonstration of the usefulness of a real biomimetic sensor based on
molecular imprints is an optical fiber-like device in which a fluorescent amino
acid derivative (dansyl-L-phenylalanine)binds to the polymer particles, resulting
Table 8 ExamplesofBiolnilneticSensorsBasedonMolecularlyImprintedPolymers
(MIPs)
Range
Analyte
Morphine
Vitamin K,
Phenylalanine anilide
Dansyl-L-phenylalanine
Atrazine
Conductometry
Sialic acid
(pg/mL) Ref.
0- 10
0-4
Qualitative
33-3300I05
0-30
Transducer
Amperometry
Ellipsometry
Capacitance
Potentiometry
fluorescence
Fiberoptic
0-0.5
0-3
fluorescence
Optical
102
103
I04
106
I07
I08
Deshpande
370
in fluorescent signals that vary as a function of the concentrationof the derivative

(106). Chiral selectivity was shown by using the corresponding D-enantiomer as
a control.
A major limitation of using MIP-based biomimetic sensors is the long response time (15-60 min) for the measurement of the desired analyte (99). This
delay could be minimized by optimizing the kinetics and selectivityof polymers.
Generally, the use of highly rigid polymers
favors selectivity but increases responsetimes.Similarly,polymerporosityincreasespolymer-bindingcapacity
and response time. Using small polymer particles or thin polymer layers should
improve diffusion rates and thus the apparent binding kinetics, giving far lower
response times. Alternatively, the initial binding (preequilibrium) rate could be
used to determine analyte concentration.
The current generationof MIP biomimetic sensors is 100- to 1000-fold less
sensitive than other types of biosensors. Although further improvements in MIPS
are likely to decrease the sensitivity gap and lead to useful applications, biomimetic and other types of biosensors will most probably find their own niches in
the future.
XI.
FUTURE TRENDS
The ability of biosensors to provide real-time information and their relatively

simple design should result in a wide variety of applications for monitoring the
quality and safety of our food supply. Indeed, sensors basedon ATP bioluminescence and electronic noses, to some extent, are already being used extensively in
the food industry. Amperometric and potentiometric biosensors have also found
widespread acceptance for measuring a range of food analytes. Optical biosensors
will find potential applications for measuring and monitoring concentrations of
such compounds as acetaldehyde, alanine, malate, lactate, nitrate, glucose, glycerol, ethanol, xylitol, isocitrate, glutamate, sorbitol, and galactose in a variety of
food processing operations; for detecting food contaminants such as Salrnotzella,
E. coli, Listeria, and Staphylococcus; and for monitoring contamination of meat
and dairy products by residues of antibiotics, growth hormones, and veterinary
drugs.
Thermal transducer-based biosensors have already been used successfully
formonitoringseveralcompoundsrelevant
to foodqualitycontrol,including
ascorbic acid, glucose, lactose, triglycerides. cholesterol, galactose, ethanol, sucrose, antibiotic and therapeutic agents, and oxalic acid. These biosensors could
also be applied to detect and monitor microbial contamination of food products.
Althoughthebiosensorsdescribedabovehavebeenimprovedtremendously, they have not yet established themselves as providing a technology that
is cheap and versatile. Immunosensing is also unlikely to be achieved as easily or
Biosensors
371
as cheaply as with the other typesof biosensors described. Among the biosensor
technologies described here, the mass-detection piezoelectric sensorsin the form
of acoustic wave guide devices, or one of the four or five optical devices described, are the most likely candidates as general purpose bioanalytical sensors
to produce and possess a potential for use over a
of the future. They are easy
wide rangeof analyte concentrations. They also approach the low detection limits
required for immunosensing, and the reproducibility of results is similar to that
achieved by conventional analytical methods. These devices rely on relatively
well-understood physical principles, and the associated instrumentation
is quite
in the optical sensor is, however,
easy to construct. The analyte detection element
much cheaper to construct and is more flexible in the uses to which it can be put
than the piezoelectric devices.
The fluorescent methods used with the optical sensors, although intrinsically more sensitive than the other optical methods, suffer from potentially greater
interference effects arising from the
test specimen matrix. It is for this reason
that the surface plasmon resonance (SPR) or integrated optical sensors are likely
to be commercialized as the next generation of general purpose analytical tools.
These devices have not yet met all the criteria listed for the ideal biosensor,
but they have met most of these criteria,
and future developments are likely to
enable all of them to be met within the next few years.
Although biosensors hold considerable promise and potential for on-line
still in its
quality and safety monitoring in the food industry, the technology is
of biosensors have been developed commercially.
infancy. Only a limited number
The problems associated with reliable mass production and commercialization
of this technology are enormous. They include long-term stability
of enzymes,
antibodies, and bioreceptors; sterilization of sensors; nonspecific adsorption of
other species; immobilization and mediation of enzyme-based sensors; reduction
of interferences from other substancesin the food; miniaturization of sensors and
sensor arrays; and variability in large-scale manufacturing. Progress on all these
fronts is continually being made in this rapidly growing area of research.
of fields involved in the developmentof biosenBecause of the wide variety
sors, successful research and developmentof biosensors will require a multidisciplinary approach. Biologists, chemists, physicists, and electrical and food and
bioprocess engineers will have to join together to solve the technical problems
facing biosensors. Nevertheless, it will not be too long before this technology
will become the mainstay of quality and safety monitoring in the food industry.
REFERENCES
I.
CDE Todd. Preliminary estimates ofcost of foodborne diseases in the United States.
J Food Prot 52:595-601, 1989.
372
Deshpande
2. HN Bean, MP Griffin.Foodbornediseaseoutbreaks
in theUnitedStates19731987: Pathogens, vehicles, and trend. J Food Prot 53:711-728, 1990.
3. MP Doyle. Food safety in the 21st century. Dairy Food Environ Sanit 13:490-493,
1993.
4. DYC Fung. Whats needed in rapid detection of foodborne pathogens. Food Techno1 49(6):64-67, 1995.
5. BJ Schaertel,RFirstenberg-Eden.Biosensorsinthefoodindustry.Presentand
future. J Food Prot 51:811-820, 1988.
6.ANJina,MJ
Tierney.Biosensorsprinciplesandapplications.In:PSingh,
BP
Sharma, P Tyle, eds. Diagnostics in the Year 2000. New Van
York:Nostrand Reinhold 1993, pp 343-366.
7. SS Deshpande, RM Rocco. Biosensors and their potential use in food quality. Food
Techno148(6):146-150,1994.
8. SS Deshpande, BP Sharma. Diagnostics: Immunoassays, nucleic acid probes, and
biosensors. Two decades of development, current status, and future projections in
clinical, environmental, and agricultural applications. In: P Singh, BP Sharma, P
Tyle, eds. Diagnostics in the Year 2000. New York: Van Nostrand Reinhold, 1993,
pp 459-525.
9. MAlvarez-Icaza,UBilitewski.
Mass production of biosensors.AnalChem65:
525A-533A,1993.
10. DG Buerk. Biosensors. Theory and Applications. Lancaster, PA: technomic, 1993.
1 I . PGMalan.Immunologicalbiosensors.In:DWild,
ed. TheImmunoassayHandbook. New York: Stockton Press, 1994, pp 125-134.
12. BM Paddle. Biosensors for chemical and biological agents of defense interest. Biosensors Bioelectron 1 1:1079- 1 1 13, 1996.
13. M Madou, MJ Tierney. Required technology breakthroughs to assume widely accepted biosensors. Appl Biochem Biotechnol 41:109-115, 1995.
14. JS Schultz, S Mansouri, I Goldstein. Affinity sensor. A new technique for developing implantable sensors for glucose and other metabolites. Diabetes Care 5:245253,1992.
15. B Liedberg, C Nylander,I Lundstrom. Surface plasmon resonance for gas detection
and biosensing. Sensors Actuators 4:299-304, 1983.
16. I Pockrand, JD Swalen, R Santo,ABrillante, MR Philpott.Opticalpropertiesof
organic dye monolayers by surface plasmon spectroscopy. J Chem Phys 69:400 I 401 I , 1978.
17. MT Flanagan, RH Pantell. Surface plasmon resonance and immunosensors. Elect
Lett 20:968-970, 1989.
18.RMPettigrew.AssayTechnique.InternationalPatentApplication.IntlPub1No.
W084/02578,1984.
19. E Kress-Rogers, APFTurner. Immunosensors based on acoustic, optical and bioelectrochemical devices and techniques. In:BA Morris, MN Clifford, R Jackman, eds.
Immunoassays for Veterinary and Food Analysis-I. Barking: Elsevier, 1988, pp
227-253.
20. GA Robinson. Optical immunosensing systems: meeting the market needs. Biosensors Bioelectron 6:183-191, 1991.
21. TE Hirschfield. Fluorescent immunoassay employing optical fiberin capillary tube;
Biosensors
antigen-antibodycompleximmobilizedonglasssurface.
373
U.S. Patent4,447,546
( 1984).
22. MD Ward. D Buttry. In situ interfacial mass detection with piezoelectric biosensors.
Science 249: 1000- 1007, 1990.
23. T Nomura, M Okuhara. Frequency shifts of piezoelectric quartz crystals immersed
in organic liquids. Anal Chim Acta 143:237-241, 1982.
24. Microgravimetric assay to be marketed. Chem Eng News (Sept. 9):23, 1985.
25. J Janata, GF Blackburn. Immunochemical potentiometric sensors. Ann NY Acad
Sci 428:286-292, 1984.
26. M Meyerhoff, GA Rechnitz. Electrode-based enzyme immunoassay using urease
conjugates. Anal Biochem 95:483-493, 1979.
of calorimetricsensors.In:
27. BDanielsson,KMosbach.Theoryandapplications
APF Turner, I Karube, GS Wilson, eds. Biosensors: Fundamentals and Applications. Oxford: oxford University Press, 1987, pp 575-596.
28. K Mosbach, B Danielsson. Thermal bioanalysers in flow streams: enzyme thermistor devices. Anal Chem 53:83A-94A, 1981.
29. SL Brooks, APF Turner. Biosensors for measurement and control. Measure Control
20~37-43, 1987.
30. J Janata. An immunoelectrode. J An1 Chem SOC 97:2914-2916, 1975.
of antigen31. N Yamanloto, Y Nagasawa, M Sawai. Potentiometric investigations
antibody and enzyme-enzyme inhibitor reactions using chemically modified metal
electrodes. J lmmunol Methods 22:309-3 17, 1978.
32. MA Arnold. An introduction to biocatalytic membrane electrodes.Am Lab 15:3440,1983.
33. MYK Ho, GA Rechnitz. An introduction to biosensors. In:RM Nakamura, Y Kasahara, GA Rechnitz, eds. Immunochemical Assays and Biosensor Technology for
the 1990s. Washington, DC: American Society Microbiology, 1992, pp 275-290.
34. I Deng, C Enke. Adenosine-selective electrode. Anal Chem 52: 1937- 1940, 1980.
35. CP Pau, G Rechnitz. Bound cofactor/dual enzyme electrode system for a alanine.
Anal Chim Acta 160:141-147, 1984.
36. MA Arnold, GA Rechnitz. Comparison of bacterial, mitochondrial, tissue, and enzyme biocatalysts for glutamine selective membrane electrodes. Anal Chem 52:
1170-1 174, 1980.
37. DP Nikolelis, DS Papastathopoulos, TP Hadjiioannou. Construction of a guanine
enzyme electrode and determination of guanase in humanbloodserumwithan
1981.
ammonia gas sensor. Anal Chim Acta 126:43-50,
38. MA Jensen, GA Rechnitz. Enzyme sequence electrode for D-gluconate. J Membr
Sci 5:117-127, 1979.
39. PM Kovach, ME Meyeroff. Development and application of a histidine-selective
biomembrane electrode. Anal Chem 54:2 17-220, 1982.
for the determination of salicylate. Anal
40. T Fonong, GA Rechnitz. Enzyme electrode
Chim Acta 158:357-362. 1984.
41. CG Guilbault, FR Shu. Enzyme electrodes based on the use of
a carbon dioxide
sensor. Urea and L-tyrosine. Anal Chem 44:2161-2166, 1972.
42. GA Rechnitz, RK Kobos, SJ Riechel, CR Gebauer. A bioselective membrane electrode prepared with living bacterial cells. Anal Chim Acta 94:357-365, 1977.
374
Deshpande
43. RK Kobos, GA Rechnitz. Regenerable bacterial membrane electrode for

l.-aspartate.AnalLett10:751-758,1977.
44. MA Jensen, GA Rechnitz. Bacterial metnbrane electrodefor I.-cysteine. Anal Chiln
Acta101:125-130,1978.
45. M Hikuma, H Obana, T Yasuda, I Karube, S Suzuki. A potentiometric microbial
sensor based on immobilized Escherichia coli for glutamic acid. Anal Chim Acta
116:61-67, 1980.
46. GA Rechnitz, TL Riechel, RK Kobos, ME Meyerhoff. Glutamine-selective membrane electrode that uses living bacterial cells. Science 199:440-441, 1978.
47. RR Walters, BE Moriarty, RP Buck. Pseudomonas bacterial electrode for determination of L-histidine. Anal Chem 52: 1680- 1684, 1980.
48. TL Riechel, GA Rechnitz. Hybrid bacterial and enzyme membrane electrode with
nicotinamide adenine dinucleotide response. J Membr Sci 4:243-250, 1978.
49. RK Kobos, DJ Rice, DS Flournoy. Bacterial membrane electrodefor the determination of nitrate. Anal Chem 51:1122-1125, 1979.
50. RK Kobos, HY Pyon. Applicationof microbial cells as multistep catalysts in potentiometric biosensing electrodes. Biotechnol Bioeng 23:627-633, 198 1.
51. CL Dipaolantonio, GA Rechnitz. Stabilized bacteria-based potentiometric electrode
for pyruvate. Anal Chim Acta 148: 1-12, 1983.
52. CL Dipaolantonio, MA Arnold,GARechnitz.Serine-selectivemembraneprobe
based on immobilized anaerobic bacteria and a potentiometric ammonia gas sensor.
AnalChimActa128:121-127,1981.
53. RK Kobos. Preliminary studies of a bacterial sulfate electrode. Anal Lett 19353362, 1986.
54. SR Grobler, GA Rechnitz. Determination of D(+) glucose, D(+) mannose, D(+)
galactose and D(-) fructose in a mixture of hexoses and pentoses by use of dental
plaque coupled with a glass electrode. Talanta 27:283-285, 1980.
55. CL Dipaolantonio, GA Rechnitz. Induced bacterial electrode for the potentiometric
measurement of tyrosine. Anal Chim Acta 141: 1-13, 1982.
56. T Kawashima, K Tomida, N Tominaga, T Kobayashi, HA Onishi. Microbial sensor
for uric acid. Chem Lett (Chem Soc Jpn) 5:653-656, 1984.
57. BJ Vincke, MJ Devleeschouwer, GJ Patriarche. Determination de Iacide L-ascorbique a Iaide delectrodes bacteriennes, tissulaires et enzymatiques. Anal Lett 18:
1593-1606,1985.
58. N Srnit, GA Rechnitz. Leaf based biocatalytic membrane electrodes. Biotechnol
Lett 6:209-2 14, 1984.
59. J Sidwell, GA Rechnitz. Bananatrode-an electrochemical biosensor for dopamine. Biotechnol Lett 7:419-422, 1985.
60. S Kuriyarna, GA Rechnitz. Plant tissue-based bioselective membrane electrode of
glutamate.AnalChimActa131:91-96,
1981.
61. F Schubert, R Renneberg. F W Scheller, L Kirstein. Plant tissue hybrid electrode
for determination of phosphate and fluoride. Anal Chem 56: 1677- 1682, 1984.
62. S Kuriyama, MA Arnold, GA Rechnitz. Improved membrane electrode using plant
tissue as biocatalyst. J Membr Sci 12:269-278, 1983.
as biocatalyst forurea biosensors. Biotech63. MA Arnold, SA Glazier. Jack bean meal
no1 Lett 6:313-318, 1984.
Biosensors
375
64. S Uchiyarna, GA Rechnitz. Biosensors using flower petal structures. J Electroanal

Chem222:343-346,1987.
65. F Schubert, U Wollenberger. F Scheller. Plant tissue-based anlperometric tyrosine
electrode. Biotechnol Lett 5:239-242, 1983.
66. D Wijesuriya, GA Rechnitz. Mixed carbon paste-pea seedling electrochemical sensor for measuring plant growth-regulating activity of amines. Anal Chim Acta 243:
1-8,1991.
67. MA Amold, GA Rechnitz. Tissue-based membrane electrode with high biocatalytic
activity for measurement of adenosine 5-monophosphate. Anal Chem 53: 18371842,1981.
68. MA Arnold, GA Rechnitz. Optimization of a tissue-based membrane electrode for
guanine. Anal Chem 54:777-782, 1982.
69. YL Ma, GA Rechnitz. Porcine kidney tissue based membrane electrodefor glucosamine-6-phosphate. Anal Lett 18: 1635- 1646, 1985.
70. GA Rechnitz, MA Arnold, ME Meyerhoff. Bioselective membrane electrode using
tissue slices. Nature 278:466-467, 1979.
RM Nakamura, Y Kasahara,GA
71. AM Yacynych.Amperometricbiosensors.In:
Rechnitz, eds. Immunochemical Assays and Biosensor Technology for the 1990s.
Washington, DC: Amer. Soc. Microbiol.. 1992, pp 335-354.
72. S Suzuki, F Takahashi, 1 Satoh, N Sonobe. Ethanol and lactic acid sensors using
electrodes coated with dehydrogenase-collagen membranes. Bull Chem Soc Jpn
48:3246-3249,1975.
73. GFH Blackburn, J Shah, J Kenten. Electrochemiluminescence detection for developmentofimmunoassaysandDNAprobeassaysforclinicaldiagnostics.Clin
Chem 37: 1554- 1539, I99 1.
74. EW Chappelle, GV Levin. Use of the firefly bioluminescence reaction for the rapid
detection and counting of bacteria. Biochem Med 2:41-52, 1968.
75. PE Stanley. A concise beginners guide to rapid microbiology using adenosine triphosphate (ATP) and luminescence. 1n:PE Stanley,BJ McCarthey, R Smither, eds.
ATP Luminescence: Rapid Methods in Microbiology. 0xford:blackwell Scientific,
1989, pp 1-11.
76. LJ Kricka. Ultrasensitive chemi- and bioluminescent methods. Prospects in food
testing. Food Testing Anal 3(4):20-23, 1997.
77. A N Sharpe, MN Woodrow, AK Jackson. Adenosine triphosphate (ATP) levels in
food contaminated by bacteria. J Appl Bacteriol 33:758-767, 1970.
78. MW Griffiths. The role of ATP bioluminescence in the food industry: new light
on old problems. Food Techno1 50(6):62-72, 1996.
79. AL Kyriakides, PD Patel. Luminescence techniques for microbiological analysis
of
foods. In:P Patel, ed. Rapid Analysis Techniques in Food Microbiology. Glasgow:
Blackie Academic and Professional, 1994, pp 196-231.
80. PC Vasavada. Advancesin pathogen detection. Food Testing Anal 3(2): 18-23, 47,
1997.
81. KJ Strassburger. Electronic nose evaluation in theflavorindustry.FoodTesting
Anal2(5):22-27,37,1996.
82. DG Moulton, Y Zotterman. Olfactory and Taste. London: Pergamon Press, 1963,
pp 7 1-97,
376
Deshpande
83. JWGardner, PN Bartlett.SensorsandSensorySystemsforanElectronicNose.
Dordrecht, The Netherlands: Kluwer, 1992, pp 161 - 179.

EKress-Rogers.HandbookofBiosensorsandElectronicNoses:Medicine,Food
and the Environment. Boca Raton, FL: CRC Press, 1996.
85. K Persaud, G Dodd. Analysis of the mammalian olfactory system using
a model
nose. Nature 29932-355, 1982.
86. K Persaud. P Pelosi. An approach to an electronic nose. Trans Am Soc Artif Intern
Organs 31 :297-300, 1985.
87. T Oishi, M Kaneyasu, A Ikegami. Discriminationof chemical compounds and functional groups by pattern recognition using an integrated sensor. Hybrid Circuits16:
19-22,1988.
88. H Abe, S Kanaya, Y Takahashi, S Sasaki.Combinedsemiconductorgassensor
system for detection of specific substances in particular environments. Anal Chim
Acta 219213-222. 1989.
89.UWeimar,KDSchierbaum,
W Gopel,RKowalkowski,R.Patternrecognition
methods for gas mixture analysis. Application to sensor array based upon SnO?.
Sensors Actuators B 1:93-96, 1990.
90. SL Rose-Pehrsson, JW Grate, DS Ballantine, PC Jurs. Detection of hazardous vapors including mixtures using pattern recognition analysis of responses from surface acoustic wave devices. Anal Chem 60:2801-2811, 1988.
91. K Ema,MYokoyama, T Nakamoto, T Moriizumi.Odor-sensingsystemusinga
quartz-resonator sensora m y and neural network pattern recognition. Sensors Actuators18:291-296,1989.
92. T Aishima. Aroma discrimination by pattern recognition analysis of responses from
semiconductor gas sensor array. J Agric Food Chem 39:752-756, 1991.
93.TAishima.Discriminationofliqueuraromas
by patternrecognitionanalysis of
responses from a gas sensor array. Anal Chim Acta 243:293-300, 1991.
94. D Kohl. Fundamentals and recent developments of homogeneous semiconducting
sensors. In: JW Gardner, PN Bartlett, eds. Sensors and Sensory Systems
for an
Electronic Nose Dordrecht, The Netherlands: Kluwer, 1992, pp 53-76.
95. PN Bartlett, JM Elliott, JW Gardner. Electronic noses and their application in the
food industry. Food Techno1 5 1( 12):44-48.
96. RJ Ansell, D Kriz, K Mosbach, K. Molecularly imprinted polymers for bioanalysis:
chromatography, binding assays and biomimetic sensors. Cum Opin Biotechnol 7:
89-94.1996.
97. JS Lindsey. Self-assembly in synthetic routes to molecular devices. Biological principles and chemical perspectives. A review. N
J Chem 15: 153-180, 1991.
98. K Mosbach, 0 Ramstrom.Theemergingtechniqueofmolecularimprintingand
its future impact on biotechnology. Biotechnology 14: 163-170, 1996.
99. D Kriz, 0 Ramstrom. K Mosbach. Molecular imprinting. New possibilities for sensor technology. Anal Chem 69:345A-349A, 1997.
100. G Vlatakis, LI Anderson, R Muller, K Mosbach, K. Drug assays using antibody
mimics made by molecular imprinting. Nature 361:645-647, 1993.
101. LI Anderson, R Muller, G Vlatakis, K Mosbach.Mimicsofthebindingsitesof
opioid receptors obtained by molecular imprinting of enkephalin and morphine.
Proc Natl Acad Sci USA 92:4788-4792. 1995.
84.
Biosensors
377
102. D Kriz, K Mosbach. Competitive amperometric morphine sensor based on an agar-
103.
104.
105.
106.
107.
108.
ose-immobilizedmolecularlyimprintedpolymer.AnalChimActa300:71-75,
1995.
LI Andersson,CMandenius, K Mosbach.Studiesonguestselectivemolecular
recognition on an octadecylsilylated silicon surface using ellipsometry. Tetrahed
Lett 29:5437-5440, 1988.
E Hedborg, F Winquist, I Lundstrom, LI Andersson, K Mosbach. Some studies on
molecularly-imprintedpolymermembranesincombinationwithfield-effectdevices. Sensors Actuators 37:796-799, 1993.
LI Andersson, A Miyabayashi, DJ OShannesy, K Mosbach. Enantiomeric resolution of amino acid derivatives on molecularly imprinted polymers as monitored by
potentiometric measurements. J Chromatogr 5 16:323-33 I , 1990.
D Kriz, 0 Ramstrom, A Svensson, K Mosbach. A biomimetic sensor based on
a
molecularly imprinted polymeras a recognition element combined with fiber-optic
detection. Anal Chem 67:2142-2144, 1995.
SA Piletsky, EV Piletskaya, AV Elgersma, K Yano, I Karube, YP Parhometz, AV
Elskaya. Atrazine sensing by molecularly imprinted membranes. Biosensors Bioelectron10:959-964,1995.
SA Piletsky, EV Piletskaya, K Yano, A Kugimiya, AV Elgersma, R Levi, U Kahlow, T Takeuchi, I Karube. A biomimetic receptor system for sialic acid based on
molecular imprinting. Anal Lett 29: 157-170, 1996.
10
New Techniques for Food Quality
Data Analysis and Control
Jinglu Tan
University of Missouri, Columbia, Missouri
1.
INTRODUCTION
Food quality data are gathered for various purposes and subjected to different
analyses. Instrumental food quality data are often analyzed and used for two
important objectives. First, to establish the usefulness
of an instrumental measurement as a food quality characteristic,it usually needs to be compared with human
sensory responses, which are the primary standard for food quality measurement.
Instrumental measurements are thus used as direct or indirect indicators of sensory quality attributes. Second, instrumental measurements are frequently used
for quality control purposes because instrumental techniques can be noninvasive,
rapid, convenient, and implemented on-line.
of variance and deterStatistical techniques are commonly used for analysis
mination of functional relationships. The routinely used techniques include correlationanalysis,analysis of variance(ANOVA),regressionanalysis,principal
component analysis (PCA), andpartial least squares (PLS) analysis. The conventional statistical methods are well known and can be found
in many textbooks.
In this chapter, we describe some recent developments
in food quality data analysis and food quality control.
First, we present a fuzzy set-based paradigm for
sensory data interpretation, which is followed by a neural network approach to
establishingrelationshipsbetweeninstrumentalmeasurementsandsensory
responses. Then we discuss the use of neural networks as recognizers of patterns
in quality data for the purpose of statistical process control (SPC). Finally, we
in real-time statisdemonstrate the use of instrumental food quality measurements
tical process control.
379
Tan
380
II. FUZZY SETS ANDNEURALNETWORKS

Since instrumental food quality measurements are frequently comparedwith human sensory evaluations, proper methods for interpreting sensory data and effective techniques for establishing functional relationships are extremely important
for food quality data analysis.In this section,we point out some problems associated with conventional techniques and introduce a new approach based on fuzzy
sets and neural networks ( I ) .
Sensory analysis involves evoking, measuring, analyzing, and interpreting
human judge responses to product sample characteristicsby the senses of sight,
smell, taste, touch, and hearing. A variety of scales are used to measure sensory
responses, and the choice of scales depends on the objectives of evaluation. Stevens ( 2 ) classified sensory scales into the following four categories:
Nominal scales for classification or naming
Ordinal scales for ordering or ranking
Interval scales for measuring magnitudes
Ratio scales for measuring magnitudes
In a nominal scale, sensory attributes are classified into categories that are
or magnitude relationship. Classifydifferent but do not have any particular order
ing ice creams according to their flavors is an example application of nominal
scale. An ordinal scale uses categoriesor levels ordered from low to high,
least to most, etc.; however, no assumptions are made regarding the interval size between two levels. An example ordinal scale would be the word sequence none, slight, moderate, strong, and extreme, used for assessing the
intensity of an attribute. Interval and ratio scales are like ordinal scales except
that the intervals between two adjacent levels are assumed equal-equal increment for interval scales and equal ratio of adjacent levels for ratio scales. The
widely used nine-point hedonic scale from extremely dislike to extremely
like is considered an interval scale. The popular graphical line scale with two
a type
interval
anchors of extreme marked near the line ends is also considered of
scale. Expressing the perceived intensity of a stimulus in numerical values is an
application of ratio scale because sensory responses are roughly proportional to
the stimulus intensity ratio (3).
One distinctive characteristic of sensory responses is that they are vague
andimprecise; Le., they arefuzzy. For example, we cannotpreciselydefine
moderate and strong and exactly tellthe difference between the two in
quantitative terms. A sensory score of, say, 5 is not intended to mean exactly 5.
As a result, the levels in a sensory scale, denoted with either words or numbers,
are always symbolic characterizations
of sensory responses ratherthan numerical
measures of the response magnitude. These levels are often assigned numerical
values and treated as numbers for the convenience of statistical analysis, but
New Techniques
381
they do not possess the arithmetical meanings of numerical values. Obviously,

assigning numerical values to nominal and ordinal scales is difficult to justify
because a logic relationship does not exist for the categories in a nominal scale
and the spacings between the levels in an ordinal scale are unknown. Although
as true
equal spacing is always assumed for interval and ratio scales and taken
a soundbasis.
in practice,thisassumption is notreadilyverifiableandlacks
Because of the fuzzy nature of sensory responses, the scales cannot be defined
in a space of instrumentally measurable variables. Any attachment of numerical
values to a scale will automatically imply some assumed logic relationship and
spacing of the levels, which are not necessarily reasonable and are difficult to
validate.
Fuzzy set and neural network techniques are very applicable tools for overcoming some of the difficulties in sensory analysis. A fuzzy set is defined over
some domain of interest called the universe.
The degree to which elements of
the universe belong to a fuzzy set is specified by a membership function. The
membership function values are drawn from the interval 0 1 (in mathematical
notation, from interval [0, I]). A membership grade of 0 indicates a nonmember,
I indicates a full member, and other values indicate partial members of a fuzzy
set. Thisis in contrast to the conventional crisp sets where the membership grades
are either 0 1 (in mathematical notation, drawn from set { 0, 1)), meaning that
an element in the universe is either a nonmember or full member of a crisp set.
The existence of partial memberships in fuzzy sets results in fuzzy boundaries
between two sets; in other words, an element
in the boundary area may be a
partial member of more than one fuzzy set. When a fuzzy set is defined over a
universe of real numbers, it is called a fuzzy number. A variable whose states
are represented with fuzzy sets or fuzzy numbers is a fuzzy variable. A group
of fuzzy sets may be used to represent linguistic concepts such as low, medium, and high, which are referred to as linguistic values. A fuzzy variable
with linguistic values is a linguistic variable. Detailed coverages of such basic
concepts can be found in many textbooks (e.g., Ref. 4).
Although primarily motivated by the imprecise nature of human linguistic
descriptions and reasoning (5), fuzzy set and fuzzy logic techniques have found
extremely rare applications in sensory analysis. Lincklaen Westenberg et al. (6)
used twoquestionnaires to evaluate 15 qualityattributes of fourtypes of fat
spread. In addition to the conventional graphical line scale, they asked the panelists to respond with true, borderline, or false to statements suchas The
product is hard. The panel responses to the statements were tallied
in the formof
fractions (or percentages) and usedas membership grades to classify the samples
through a fuzzy logic inference procedure. Different inference rules and attributes
were tried. They concluded that the approach was promising and deserved further
development.
An artificial neural network relates one set of variables to another through
Tan
382
interconnecting neurons, eachof which represents a linearor nonlinear functional

relationship. A neural network thus forms a complex, nonlinear model structure
for two setsof variables. The complexity of such a structure and the many adjustable parameters involved make a neural network extremely powerful for establishingfunctionalrelationships.Comparedwithconventionalregressiontechniques,
neural
networks
have
distinct
advantages
in modeling
complex
relationships between multi-input and multi-output data. This makes neural networks a tool of great potential for establishing relationships between instrumental
quality measurements and sensory responses.
A.
Fuzzy Set Interpretations of Sensory Scales, Attributes,

and Responses
By and large, sensory analysis is about modeling human senses and using human
senses as sensors for measurements. Human senses are sophisticated and powerful, and yet they are inaccurate
and lack repeatability. Exact quantification of
sensory responses is difficult if not impossible because they are naturally fuzzy.
As a result, sensory scales and attributes are also fuzzy. The following are fuzzy
set interpretations of some commonly used sensory scales, sensory attributes, and
sensory responses.
1. Sensory Scales as Fuzzy Sets

A sensory scale is usually a set of terms that describe the intensity levels or the
categories of some stimulus attributesof interest. The attributes and their intensity
ranges or categories specify a space or universe
in which the set of terms are
defined. For example, sweetness in a range may be the attribute of interest and
evaluated with a scale consisting of the term set (slight, moderate, very much).
The range of sweetness is the universe in which the terms are defined. In reality,
however, the terms in a sensory scale are only vaguely described. Each term may
represent a stimulus intensity in a general neighborhood, and there is not a clear
fit the
boundary between two adjacent terms. Such linguistic terms obviously
to
description of fuzzy sets. In fact, they are exactly what fuzzy sets are used
represent. A sensory scale is simply a battery of fuzzy sets.
a. Notninal and Ordinal Scales. In nominal and ordinal scales, a sample

is classified into a category to which the sample is closest. Once classified, the
sample is considered 100% belonging to that category; i.e., it has a membership
grade of 1 in that set. This may appear to be a caseof the conventional crisp set,
in which a sample either fully belongs to a set or it does not. A crisp set has a
crisp boundary, B, in the universe, x, between two adjacent sets, C, and C,+l(Fig.
I). The membership function, p, is a constant 1 over the domain of each set. TO
the left of B, a sample is classified as C, and otherwise C,,,. In sensory scales,
383
New Techniques
Fig. 1 Two crisp sets, C, and C,,,, a crisp boundary, B, in universe, x.
however, a crisp boundary B never exists. As a result, an alternative and more

reasonable representation of nominal and ordinal scales is a battery of singleton
fuzzy sets (Fig. 2 ) . A singleton fuzzy set is one which has only one member.
This member has a membership grade of 1 in the set. This agrees with the fact
that the categories in nominal and ordinal scales are considered mutually exclusive and a member in a category (set) is always a full member. We may further
consider that each singleton set is defined over a general vicinity of a point in
the universe rather than precisely at the point. In other words, if a sample is near
a singleton set, it is classified as a full member of the set. The fuzziness of the
word near implies that there is not a clear-cut boundary between two sets.
Furthermore, the singleton sets are ranked
in an ordinal scale and are not
in a
nominal scale.
0. Interval and Ratio Scales. Among the interval and ratio scales in use
today, variations exist. Some do but others do not permit simultaneous partial
classification of a sample into more than one level or, simply termed, betweenlevel classification. For interval and ratio scales thatdo not permit between-level
classification, a suitable representation would also be a batteryof singleton fuzzy
Fig. 2 A battery of singleton fuzzy sets, C , to C,, describing nominal and ordinal scales,
and interval and ratio scales that do not permit between-level classification.
384
Tan
Fig. 3 A battery of fuzzy sets, C, to C,, with triangular membership functions depicting
interval and ratio scales that permit between-level Classification.
sets (Fig. 2 ) . For example, the nine-point hedonic scale is considered an interval
scale, and a sample must be classified into one
of the nine levels. Indeed, such
a scale is not fundamentally different from an ordinal scale except that the intervals are assumed equal.
Some widely used interval or ratio scales allow between-level classification. The US. Department of Agriculture (USDA) scale for grading marbling
abundance in beef steaks is a good example. It has nine levels, from devoid
to moderately abundant. A sample may be given a grade between two adjacent
levels with the closeness to the higher level indicated by a degree. For example,
a 30-degree slight grade means that the sample stands at the 70-30% split of
the interval between traces and slight. Such a scale can be exactly represented by a battery of fuzzy sets defined by the membership diagram in Fig. 3.
which is routinely seen in the fuzzy set and fuzzy logic literature. The fuzzy sets
or linguistic values, C,, C2, etc., are defined by a set of triangular membership
functions as shown by the solid lines in Fig. 3. When the intensity of an attribute
in the universe, x, its membership grade
under evaluation moves from left to right
decreases in one set and increases in another in a linear fashion. For instance, a
sample with an attribute intensity x, partially belongs to both C,and C2 (Fig. 3 ) .
Its membership grades in the two sets are pI(x,) and p?(xl), respectively, with
the two summing to 1. Evidently the membership grades are simply a different
interpretation of the degrees in the USDA marbling scale, and the fuzzy sets
rather accurately depict the marbling abundance levels.
The intervals between
the vertices of two adjacent membership functions may not be equal in general,
but in the case of an interval scale they are often assumed equal as discussed
before.
The widely used graphical line scale for intensity evaluation
is simply a
An
special case of the interval scale that permits between-level classification.
example line scale is shown in Fig. 4a. Two intensity extremes, such as exof the line. A
tremely low and extremely high, are marked near the ends
385
New Techniques
Extremely
Low
I
Extremely
High
I
Extremely
High
Fig. 4 (a) A graphical line scale. (b) Two fuzzy sets representing the line scale.
judge indicates his or her evaluation of an attribute by putting a mark between

the two extreme tics. This
mark is thenconvertedinto a numericalscore by
measuring its distances from the extremetics. The conversion is most commonly
done according to a linear distance-to-score relationship with the low extreme
as the zero-score point and the high extreme as the full-score point. Such a scale
is apparently well described by the fuzzy sets depictedin Fig. 4b. There are only
two sets, extremely low and extremely high, each
of which is defined by
a triangular membership function. When the intensity
of a measured attribute
increases, its membership in extremely low decreases and that in extremely
high increases in a proportional manner. Of course,a nontriangular membership
function can be used to describe a nonlinear distance-to-score relationship.
Table 1 summarizes the fuzzy set representations of the commonly used
sensory scales as discussed.
2. Sensory Attributes as Fuzzy Variables

Because sensory scales are fuzzy sets used to represent the intensity values
or
states of sensory attributes, sensory attributes are fuzzy variables
by definition
(variables whose states are represented with fuzzy sets). Furthermore, since the
fuzzy sets in sensory scales usually denote some linguistic concepts
or values,
sensoryattributesare also linguisticvariables(fuzzyvariableswithlinguistic
values).
Tan
386
Table 1 SummaryofFuzzySetRepresentationsofCommonlyUsedSensoryScales
Fuzzy
Scale
Nominal scales
Ordinal scales
Interval and ratio scales
that do not permit
between-level classifications
Hedonic scales
Interval and ratio scales
that permit betweenlevel classifications
Line scales
Unranked singleton fuzzy

sets
Ranked singleton fuzzy
sets
sets
Singletons defined near a

point (see Fig. 2)
As above

sets
Regular ranked fuzzy sets
As above
Two regular fuzzy sets
As above
Other shapes of membership functions possible

(see Fig. 3 )
Other shapes of membership functions possible
(see Fig. 4)
3. Sensory Responses as Membership Grades

In sensory evaluation, the state of a sensory attribute is measured by one or more
human judges. The judges arefirst trained according to some predefined sensory
scale, which is usually presented as a set of standard samples. This is simply the
process of sensor (human senses) calibration against a standard. After training,
the judges are usedto evaluate the sensory attributeof product samples.A judges
or her measure of the state of the sensory
sensory response to a sample is his
attribute. Since a sensory attribute is a fuzzy variable, whose states are expressed
in fuzzy sets, the sensory response is simply the judges evaluation of the membership grades of the sample in the fuzzy sets of the scale used.
In all the scales that can be described by singleton fuzzy
sets (including
nominal scales, ordinal scales, and interval and ratio scales that
do not permit
between-level classification), judges are asked to assign a sample to one of the
linguistic categories represented by fuzzy sets C,, C:, . . . C,,. If judge j decides
that a sample with attribute intensity x belongs to category C,, then the sample
membership grade in fuzzy set C, given by judge j, pi,,(x),is 1; otherwise, it is
0; i.e.,
Pi,jtX) =
if C, is chosenfor x by judge j
0 otherwise
(1)
New Techniques
387
In an interval or ratio scale that permits between-level classification, a judge

may select a point between two adjacent levels, C, and C,,,(i = 1, 2, . . . , n I), for a sample with attribute intensity x. The membership grades of the sample
in C, and C,,,determined by judge j are, respectively, pl,,(x) and pltl,J(x),
which
sum to 1 ; i.e.,
Pi.J(X) +
CLi+l.J(X)
= 1
(2)
The specific values of P,,~(x) and P,+~.,(X)

depend on the membership functions.
in Figs. 3 and 4, the
In the case of triangular membership functions as shown
pI+I,J(x)vary in an inversely proportional manner between two
values of V,,~(X) and
adjacent levels. For example, a 30-degree slight grade for marbling abundance
given by a judge to a steak sample means that the sample has a membership
in fuzzy set traces and 0.3
(P,+,,~)in slight, and a 40grade of 0.7 (p,,,)
degree slight grade means a membership grade of 0.6 in traces and 0.4 in
slight.
4.
Fuzzy Membership Vector and Fuzzy Histogram of

Response
Since human sensory responses are intrinsically inaccurate and imprecise, variations occur from judge to judge and from one measurement to the next by the
same judge. These variations constitute a measurement noise, which can be reduced by calibration (judge training) but cannot be eliminated.As a result, multiin sensory
ple judges are used to minimize the effect of this noise, especially
studies where data accuracy
is essential. When multiple judges are involved, there
exists the question of how the data from multiple judges should be represented;
in other words, how can we establish an overall response for a sample?
The conventional way of finding an overall response is to treat the individual judge responses as numerical scores and take the average
of them. This suffers
from the difficulty in providing the underlying equal-interval and linearity assumptions as discussed before. Sincethe responses are values of fuzzy variables,
a logical way would be to take the fuzzy average of them. However, the fuzzy
operations based on the extension principle (7) and consequently the fuzzy averages (8) require explicit knowledgeof the membership functions.In sensory analysis, the fuzzy setsin the scales are not explicitly defined in terms of some instrumentallymeasurablevariables;thus,theaveragingoperationbasedonthe
extension principle is not applicable.
While there may be many possible waysto represent and utilize multijudge
responses, one method would be the fuzzy histogram or membership vector described as follows. Though a fuzzy average over the fuzzy values
is not obtainable
without explicit knowledge of the membership functions, the membership grades
given by the judges can be averaged within each fuzzy set. Suppose that there
388
Tan
0.7 1
c, c, c,
c, c, c,
Fuzzy set or category Ci
c,
Fig. 5 An example plot of fuzzy membership vector
c,
c,
or fuzzy histogram of response.
a
are n fuzzy sets (categories) in the scale used and there are m judges. For
sample with attribute x, judge( jj = 1,2, . . . m) decides that its membership grade
in fuzzy set i (i = I , 2, . . . n) is p,,,(x) as describedin the previous subsection. Note
that p),,(x)= 0 if fuzzy set C, is not chosen by judge j. The average membership
grade of the sample in fuzzy set C, voted by all the judges is
TP1.,(X)
pix) =
,=I
(i = 1 , 2 . . . n)
(3)
pi(x) represents an averageor consensus voteby all judges on the degree to which
the sample belongs to category C,; in other words, it is an expected membership
grade of x in fuzzy set C,. Note that the averaging operation is performed within
a fuzzy set and does not require knowledge of the membership functions and the
n fuzzy sets in the scale. Eq.
intervals between the fuzzy sets. Since there are
(3) implies n membership grades for x, which can be written into the following
fuzzy membership vector
P(X) = [PI(X) P?(X) . . CI,I(X)IT
'
(4)
where superscript T stands for transpose.

Fuzzy membership vector ~ ( xshows
)
the membership grades of x in all
n fuzzy sets. Graphically, it can be plotted as a bar graph exemplified by Fig. 5 .
Since Fig. 5 is actually a normalized histogram of the judge responses, which
New Techniques
389
are membership grades in the fuzzy sets of a scale, we may refer to Fig. 5 as a
fuzzy histogram of response. Because no assumptions are involved, the fuzzy
of response provides an unadulterated
membership vector or the fuzzy histogram
depiction of the overall response by a panel of judges.
It should be noted that a single judge is also used in practice because of
cost considerations as in the case of government-administered grading. A single
judge is just a special case of a judge panel with m = 1. The fuzzy membership
vector and fuzzy histogram of response defined are still applicable.
5. Defuzzification
The fuzzy membership vector or fuzzy histogram of response gives an unaltered
representation of judge responses for sensory analysis, and the fact
that a sample
may be partial membersof multiple categories depicts the fuzzy nature of sensory
responses. In commercial applications of sensory evaluation, however,it is sometimes necessary to assign a sample to a single, specific category. Then we need
to convert the multijudge responses into a unified consensus, or in other words,
to convert the fuzzy membership vector into a scalar of unity corresponding to
a single crisp set. The conversion
is similartodefuzzification in fuzzylogic
control. Several defuzzification methods have been proposed in the fuzzy logic
literature, including the maximum method, the mean-of-maximum method, and
the center-of-gravity method (9, IO).
The concepts of the defuzzification methods may be borrowed to convert
membership vectors into crisp sets. By the maximum method, the fuzzy set with
the maximum membership grade is chosen as the defuzzified crisp value (or category). Lincklaen Westenberg et al. (6) employed this method to determine the
in Fig. 5
final classification. For example, the fuzzy membership vector plotted
will be defuzzified by the maximum method to givea crisp classification of category C,. When there is a tie in the maximum membership grade, any one set
involved in the tie may be chosen as the crisp value. The mean-of-maximum and
the center-of-gravity methods use the center of peaks and the center of gravity,
respectively, of a combined membership function. This requires explicit knowledge of the membership functions. As the membership functions are generally
unknown in sensory analysis, these two methods are rarely applicable.
The maximum method is a logical choice for defuzzification of multijudge
sensory responses in many cases. The fuzzy set or category with the maximum
membership grade represents the majority opinion of the judges and thus seems
to be a natural and logical choice for the crisp value or consensus classification.
In contrast, the conventional defuzzificationby arithmetic averaging of assigned
numerical values is not as natural because an unverifiable equal-interval assumption must be made, as discussed before, and also because the resulting crisp value
does not generally possess the majority-opinion property.
Tan
390
6. Neural Network Prediction of Sensory Responses From
Instrumental Measurements
Sensory responses are reactions of senses to stimuli. Stimuli can be measured
and characterized by instrumental measurements.
It is of great interest to researchers and product developers to establish relationships between sensory responses and instrumental measurements of stimuli. Such relationships can lead
to a fundamental understanding of human senses and help formulate products
that people like. Moreover, they establish the usefulness of instrumental measurements as indicators of food quality.
As discussed in the previous section, it is unreasonable to represent sensory
responses to a product sample as a precise number.
The fuzzy membership vector,
p, or the fuzzy histogram of response describes the judges' responses without
artificial modification. One obvious characteristic of p is that it is multivalued
in general. At the same time, characterization of a stimulus or a set of stimuli
the stimuli for human responses
usually requires a multitude of data. For example,
of marbling abundance in a beef steak may include a numberof marbling characof the
teristicssuchassize,uniformity,percentarea,andspatialdistribution
marbling flecks. As a result, relationships between instrumental measurements
and sensory responses are routinely multi-input and multi-output (MIMO) type.
Moreover, the relationships areusually nonlinear because of the nonlinear nature
of human senses.
To establish nonlinear relationships for MIMO data, the conventional statistical methods are extremely awkward.Artificial neural networks, fortunately, provideaneffectivetoolforthispurpose.
An artificial neuralnetworknormally
consists of a layer of input nodes, a layerof output nodes, and oneor more layers
o f hidden nodes in between. Between every two adjacent layers, the nodes are
interconnected with neurons, each of which represents some linear or nonlinear
relationship with two adjustable parameters: a weight and a
bias. Overall, each
output node is related to the input nodes through a complex nonlinear relationship. Many algorithms and techniques are available to determine the parameters
for an optimal fit of a functional relationship between the inputs and outputs.
to many available text(Readers unfamiliar with neural networks are referred
books, including Ref. 1 I.) In mathematical notation, an artificial neural network
represents the following nonlinear relationship, f
f(X)
(5)
with
x = [x, X? . . . XJT
[ Y I Y? . . . Y"1'
where X and Y are the input and output vectors, respectively, and k and
the numbers of inputs and outputs, respectively.
n are
New Techniques
391
For sensory response prediction,Y will be the sensory response represented

by fuzzy membership vector p (i.e., Y = p), and X will be a stimulus vector
consisting of various measurable characteristics of the stimuli. For example, if Y
is sensoy responses of color, X should include various instrumentally measurable
color properties that can potentially influence human perceptions of color.
C.ExampleApplications
We demonstrate the use of the proposed fuzzy set and neural network paradigm
withtwo example applications in beef quality evaluation. The current USDA
beef quality grading system is heavily based on visual appraisals of muscle color
and marbling abundance over a steak. It is therefore of interest to study human
of various
sensory response of muscle color and marbling abundance as functions
instrumentally quantifiable characteristicsof color and marbling.In these applications, color and marbling were characterized by using color image processing
(12).
1.
Samples
Forty-three wholesale beef carcasses were used. Slices

of the small end 5 cm
thick were cut. Each slice was cutinto two 2.5 cm thick samples. The two freshly
cut surfaces, which were congruent mirror images of each other, were used for
quality evaluation: one for sensory analysis and the other for image processing.
2. SensoryEvaluation
A IO-member panel of 7 males and 3 females was assembled to evaluate beef
steak muscle color and marbling abundance. All participants were screened for
formal training in fresh beef quality evaluation. The panelists were reacquainted
official
with the attributesof fresh beef qualityin training sessions with the aid of
photographical standards of color and marbling. Representative samples were
evaluated during the training sessions and group discussions helduntil the panelists gave relatively uniform scores to the training samples. The laboratory conditions for sensory evaluation were maintained in accordance with the guidelines
established by the American Meat Science Association ( 1 3).
Muscle color of each steak was scored according to the Beef Steak Color
Guide from the National Cattlemens Beef Association (formerly National Live
Stock and Meat Board, Denver, CO) in an eight-point scale:
[Very dark red. Dark red, Moderately dark red, Slightly dark red, Cherry
red, Moderately light cherry red, Very light cherry red, Bleached
red]
Marbling abundance was rated with the aid of USDA marbling score standards
in a nine-point scale:
Tan
392
(Devoid, Practically devoid, Traces, Slight, Small, Modest,
Moderate, Slightly abundant, Moderately abundant)
3. Fuzzy Membership VectorFormulation

According to the fuzzy set interpretations, the color scale consisted
of eight fuzzy
to bleachedred(denoted
as C, to CR),andthe
setsfromverydarkred
marbling scale consisted of nine fuzzy sets from devoid to moderately abundant (denoted as C, to C,). While interval equality cannot be proven, as discussed before, both the color and marbling scales are interval scales since they
are used to measure magnitudes of sensory attributes (fuzzy variables): redness
and marbling abundance. The color scale does not allow between-level classification; thus, it is represented by a battery of singleton fuzzy sets as shown in Fig.
2. The membership grades were determined from the sensory responses
by using
Eq. (1). The marbling scale isan interval scale that allows between-level classification;then it is represented by abattery of fuzzysetsdefinedbytriangular
membership functions shown in Fig. 3. As mentioned before, a sample falling
between two levels is given a grade stated as a degree of closeness to the higher
level. The sensory responses were represented as fuzzy membership grades according to Eq. ( 2 ) and the explanations following Eq. (2). The multijudge responses were then formulated as fuzzy membership vectors or fuzzy histograms
of response according to Eq. (3). Figure 6 shows the fuzzy membership vectors
of marbling for three steaks.
0.7
- 4
USample 1
Sample 2
Sample 3
i- 0.6
0.0
c,
c2 c,c,c, c, c,
c,
c9
Fuzzy set or category Ci
Fig. 6 Fuzzymembershipvectors o f marblingabundance for threesteak samples.
New Techniques
4.
393
Instrumental Measurement of Stimulus Characteristics
Color image processing was employed to characterize the stimuli instrumentally.

Image processing algorithms were developed to segment each steak image into
or ribeye muscle, the
two subimages: one contained only the longissimus dorsi
other had only the marbling flecks. From these two subimages, the color and
marbling characteristics were computed.
a. ColorStimulusCharacteristics.
Thelongissimusdorsimuscleimages were represented in the RGB (red, green, and blue) coordinates. Muscle
color was characterized by computing the mean and standard deviation of the
of six
R, G, and B functions, which led to the following color stimulus vector
elements:
where p and (J denote mean and standard deviation, respectively, and subscripts
R, G, and B denote the color functions. The means indicated the average color
properties of a longissimus dorsi muscle, and the standard deviations showed the
color variations over it.
b. Marbling Stitnulus Charucterisfics. The size and density of marbling
flecks were considered most relevant
to sensory responses of marbling abundance. Each marbling fleck was individually identified and its size (area) calcuin 0.5mm2increments into 11 size
lated. The marblingfleckswereclassified
categories: A 5 0.5, 0.5 < A 5 1, 1 < A 5 1.5 . . . 5 < A 5 6.5, and A >
6.5; where A is fleck area in mm. The number of marbling flecks in each size
in a marbling
category was counted and divided
by the ribeye area, which resulted
for all the sizecategoriesformed a marbling
density. The marbling densities
density vector as
where d, (i = I , 2 , . . . 1 1 ) is marbling density in size category i. X,,, was the

marbling stimulus vector.
5. NeuralNetworkTraining
Feedforward neural networks were trained by using the back propagation algoMath
rithm to predict the sensory color and marbling responses. Matlab (The
Works, Natick, MA) was used for neural network training and testing. The samples were randomly separated into training and testing sets.
Two networks, one
for color and the other for marbling, were trained by feeding the training data
(6) and (7)]
in a sequential and recursive manner. The stimulus vectors [Eqs.
were the inputs, and the fuzzy membership vectors [Eq.
(4)1 were the outputs.
Different numbers of layers and neurons and different transfer functions were
Tan
394
tested to minimize the sum of squared errors. The resulting neural networks had
three layers with, respectively, 33, 20, and 8 neurons for color and 20, 25, and
9 neurons for marbling. A linear transfer function was used for the
first layer
and a sigmoid function used for the second and third layers. The networks were
quite complex, consistingof many layers and neurons. This complexity was considered to be a result of the complex nature of sensory responses to stimuli.
6. SensoryResponsePrediction
After the neural networks were trained, the
test data sets were used
to verify
them. The inputs in the test data were fed to the networks to generate predicted
outputs. The predicted fuzzy membership vectors were compared with
the real
sensory fuzzy membership vectors.
Figures 7 and 8 show comparisonsof predicted and sensory fuzzy membership vectors for a test sample. Results for other test samples were similar. The
predicted and sensory fuzzy membership grades had very similar patterns. For
a test data set of five samples, the mean squared errors in the predicted membership grades averaged 0.0059 for color prediction and0.0012 for marbling prediction. This shows that the neural networks could predict the sensory responses to
a good degree of accuracy for both muscle color and marbling.
It should be noted that the network outputs were fuzzy membership vectors
Sensory
E2223 Predicted
1
Color fuzzy set or category Ci

Fig. 7 Comparison of predicted and sensory fuzzy membership vectors of muscle color
for a steak sample.
395
New Techniques
1 .o
h
i0
0.8
.f3.
0.6
xb
5
B
c,
0.4
E
0
M
0.2
e,
0.0
c,
c2
c,
c,
c5
C6
c,
c,
c,
Marbling fuzzyset or category Ci

Fig. 8 Comparison of predictedand sensory fuzzymembershipvectors
abundance for a steak sample.
of marbling
instead of simple crisp scores or grades. The predictions were for sensory reartificial simplification.This
sponses in virtuallytheiroriginalformswithout
allows analysis of sensory data with their natural fuzziness and complexity and
determination of instrumentalmeasurements as indicators of truesensoryresponses. When the maximum method was used to classify the
test samples into
specific categories, the predictions were
100%correct forboth color and marbling
of the test samples.
7. Comment
When a new approach is developed, one would naturally expect a quantitative
comparison with existing approaches in terms of prediction accuracy and so on.
Unfortunately, such a comparison cannot be easily made meaningful because the
or do not exist in
true values (true grading, true category, etc.) are unknown
sensory evaluations. This is usually the case when subjective judgments are involved. Nevertheless, differences existin the ways subjectiveor sensory information is analyzed and interpreted. Some are more logical and reasonable than others.Theparadigmandproceduresdescribedhaveapparentadvantagesover
conventional techniques in avoiding unverifiable assumptions and retaining the
a more realisintrinsic fuzzy characteristicsof sensory responses and thus provide
tic representation of reality.
Tan
396
111.
PATTERNRECOGNITION
Statistical process control (SPC) has been an important tool for quality control
in the processing industry. It has been regaining acceptance and popularity
in
recent years becauseof the increasing consumer demand for quality products and
the intensifying global economic competition.SPC is being employed as a major
quality control mechanism for more and more manufacturing processes.
It has
become an indispensable tool to the manufacturing industry for improving productivity and product quality (14).
An SPC system consists
of means for product quality data gathering, quality
data analysis to detect process abnormalities, and implementation of corrective
all product
actions when necessary(1 5). Quality variables are evaluated for either
units or product samples taken periodically. The quality variables are plotted as
functions of time on control charts such as Shewhart and cumulative sum (CUSUM) for production process monitoring (1 6). If a quality variable deviates excessively from normal values or exhibits abnormal patterns, one or more special
causes exist and appropriate corrective actions are determined and implemented.
Depending on the product and process, one or more steps
of the SPC procedure may require human involvement. These may include manual evaluation of
data characteristics and manual implementation of corrective actions. The need
for manual operations often leadsto a considerable time lag between data gathering and process correction, which can be costly. For a high-throughput process,
it is very desirable to have on-line problem detection and process control because
of lowany time delay in process adjustment can result in significant amount
grade products or wastes. This
is especially true for continuous processes that
cannot be stopped and restarted frequently. A good example
of such a process
is food extrusion, which currently relies
on constant attention of experienced
operators for control. To implement an on-line SPC system, a critical ability
is
automatic detection of abnormalities. The 30-criterion (three-standard-deviationcriterion) detects the existence of special disturbances
by performing limit checking. The abilityto identify special disturbances by recognizing unnatural patterns
in quality data is also very important in SPC applications.
Swift (1 7) proposed a dichotomous decision tree approach and used a series
of statistical hypothesis tests to identify six unnatural patterns in control charts
a highly inflated
(quality data plots). The approach was found useful but had
false rate because of the use of hypothesis tests. The capability of the pattern
recognition algorithm depends on the accuracy of the hypothesis tests.
Cheng ( 1 8) developed a process deviation reasoning system using a traditional syntactic pattern approach to recognize unnatural patterns. Specific protoin thesystem.Thesetemplatesareusertypepatterns or templatesareused
defined and could significantly affect the performance of the resulting algorithm.
In addition. the pattern-making processis tedious and time-consuming, especially
New Techniques
397
when the number of templates or patterns is large. This has impeded real-time
applications of the approach.
With the advent of computers, intelligent pattern recognition techniques
have been developing rapidly, making automated pattern recognition possible.
Hwarng ( 19) reported a neural network model to recognize six common patterns
in quality data. The model could successfully recognize computer-generated patterns. However, it requires transformation of real data into binary numbers as
input, leading to a large input layer size and making the neural network complicated and slow.In addition, the transformation inevitably changes some statistical
real
characteristics of quality data, which can make the algorithm less reliable for
applications.
In this section, a sliding-window schemeis described for developing neural
network pattern recognizers for quality data analysis. Synthesized and experimenof the approach.
tal data from food extrusion are used to show the usefulness
The recognizers are suitable for on-line applications.
A.
PatternsinFoodQualityData
A production process usually involves the operation

of a complicated system
with many possible causes for variations. Generally, fluctuations in quality data
can be classified into two categories: natural and unnatural patterns. The major
distinction between a natural pattern and an unnatural one
is that the former is
random and the data follow a normal distribution, whereas the latter has identifiable patterns and the data do not follow a normal distribution.
The Statistical Quality Control Hardbook (20) summarizes 15 unnatural
patterns in quality data. Among them, the most commonly occurring patterns
include trend, cycle, systematic, stratification, mixture, and sudden shift. Figure
9 illustrates these patterns.
A trend is defined as a continuous and gradual movement up or down. A
series of periodic upward and downward variation is a cycle pattern. If a pattern
is systematically alternating between up and down, thenit is a systematic pattern.
A stratification pattern is a series of points that hug a centerline with a few deviations from the centerline. A mixture pattern
is characterized by points tending
to fall near the upper and lower control limits with an absenceof normal fluctuations near the middle. A sudden shift pattern shows
an abrupt change in level.
Interested readers are referred to the Statisticd Qunlity Control Hmdbook (20)
for detailed descriptions of the patterns.
B. Data Formatting for Pattern Recognizer Training

For a neural network to recognize a pattern, a segment of quality data containing
the pattern must be given to the neural network. Quality data can be presented
398
Tan
Trend
Cycle
Stratification
Mixture
I
Sudden shift
Fig. 9 Six common patterns in qualitydata.
to a neural network in a sliding-window fashion. The window length w must be

long enough to contain at least one complete pattern. In other words, quality data
sequence y(t) is formatted as neural network input vectors as
[y(t - w), . . . Y(t - 11, y(t)l
?
t = w , w + 1, . . .
(8)
where t is the current time. Equation (8) simply indicates that each input vector
to the neural network is a window or segmentof a data sequence and the window
slides forward by one data point for each time increment. To train a neural net-
New Techniques
399
work to recognize N patterns, N columnvectors(onefromeachpattern)are

presented after each time incrementto the neural network as input in an arbitrarily
selected order.
An N-bit binary number can be used to represent N different patterns. For
example, if there are six patterns of interest, [ I , 0, 0, 0, 0, 0IT indicates pattern
I , 10, I , 0, 0, 0, 01 indicates pattern 2, and so on. The binary numbers are used
as outputs to train a neural network pattern recognizer.
C.PatternDataGeneration
It is generally difficult to acquire
real quality data with distinct and repeated
patterns of importance for recognizer training. As
an alternative, a pattern generator may be used. A pattern generator may be expressed in a general form of a
process mean plus two disturbances:
Y(t) = I.I + d(t) +
at)
(9)
where y(t) is a quality measure at time t, p is the process mean for y(t) when
the process is in control, d(t) is a special disturbance, andE(t) is a random variable
following a normal distribution, i.e.,
E(t) = N(O, ro)
(10)
where o is the process standard deviation when the process

is in control and r
is the magnitude of random noise in terms of CJ (0 < r < I ) .
Special disturbance d(t) determines the typeof pattern. The special disturbances for generating the six common patterns are shown
in Table 2.
D. ExampleApplications
We use two examples reported in Tan and Sun (21) to demonstrate the development and application of neural network pattern recognizers. One example was
based on synthesized pattern data, and the other on experimental data.
1. Synthesized DataExample
N. Pattern Data. Equation (9) was used to generate data of six patterns
by usingthespecialdisturbancesdescribed
in Table 2. For robustness of the
trained neural network pattern recognizer, the pattern data were generated with
different combinations of magnitudes of the special disturbances and noise levels
shown in Table 3. Symbol yi,,(t) denotes a data sequence generated with special
is nonpedisturbance magnitude i and noise levelj. Since the sudden shift pattern
riodic, it was generated in such a way that each vector of size w would contain
one shift at a sequentially different point in the window.
Tan
400
Table 2 SpecialDisturbancesforPatternDataGeneration
rameters
on
disturbance
Notes
Special
Pattern
Trend
d(t) = k(t - ~ J o
k = slope of trend
terms
in of
o
4, = initial time
Cycled(t)
= kosin[2x(t - &,)/TI
k = amplitude of cycle intermsof
o
T = cycle period
Systematic
d(t) = k(- 1)'o
k = amplitude of pattern
terms
in
of o,
k>0
Stratification
d(t)
= d
d = offset
from
process
the
mean (center
of stratification)
Mixture
d(t, r) = k(- 1)"s
r = random
number,
0
<r< 1
k = magnitude of pattern in terms of o,
k>0
If r < p. w = 0, otherwise w = 1
P = prespecified probability value which
determines the shifting between distributions
Sudden
shift
d(t,
t,) = k(-l)'o
t, = time
when
sudden
shift
occurs
s = 0 if t 2 t,, and shift upward
s = 1 if t < t,, and shift downward
k = magnitude of shift in terms of o,
k = 0 i f t < t,, k > 0 if t 2 t,
Table 3 Magnitude of Special Disturbance and Noise Levels Used for Pattern Data
Generation
j
Data sequence y,,,(t)

3
Magnitude
Pattern
2
level.
Noise
of disturbance,
k
~~
1.5
1.5
1.75
shift
0.1
2.5
Trend
Cycle
Systematic
2.5
Stratification
Mixture
Sudden
2.0
2.0
1 .5
I .5
0. I
0.3
0.5
0. I
0.1
0.0.3
I
0.6 0.4
0.3
0.3
0.3
0.2
0.5
0.5
0.5
0.5
New Techniques
401
b. RecognizerTraining.
The neuralnetworkrecognizerwasathreeAs a result
layer network with an input layer, a hidden layer, and an output layer.
of the data format, the input layer consisted of w nodes, and the output layer had
six nodes. Twenty-eight nodes wereused for the hidden layer. The transfer function of the hidden layer was a sigmoid function. The transfer function of output
layer was a linear function.
When 50 data vectors (windows of data) from a pattern were used to train
the recognizer and data from the same pattern at the same magnitude and noise
levels were used for testing, the results were excellent. When data from the same
pattern but at different magnitude and noise levels were used to
test the neural
network recognizer, the results were poor. For satisfactory robustness of the recognizer, the training data were not only from different patterns, but also from
different magnitude and noise levels. The first 300 training vectors consisted of
50 vectors from each of the six patterns at magnitude and noise levels of yz,z(t).
The next 100 training vectors were from the cycle and the mixture patterns
at
magnitude and noise levels of y,,,(t). Thelast 50 training vectors were from the
sudden shift pattern at magnitude and noise levels of y3,3(t).
The backward propagation algorithm was used to train the neural network.
Momentum and adaptive learning rate features were
used to facilitate the training
process. The Matlab Neural Network Toolbox was used for network training and
testing.
c. RecognizerPerformance.
The performance of thetrainedneuralnetwork pattern recognizer was evaluated with test data (data not used for training).
Because the recognizer might correctly or incorrectly identify a pattern or might
even fail to recognize a pattern as one of the six, three performance measures
were used.The target rate (TR) was the percentage of correctly identified patterns.
The error target rate (ETR) was the percentage of incorrectly identified patterns.
The false target rate (FTR) was the percentage of patterns unrecognized by the
recognizer(neithercorrectlynorincorrectlyidentified).Thefalsetargetrate
showed how frequently the recognizer failed to recognize a pattern.
Table 4 shows the target rate and error target rate for testing sequences of
different magnitude and noise levels (indicated by the subscripts of y; see Table
3). All target rates were more than 79% with a majority being perfect or nearly
or
perfect. All error target rates were less than 5.7%, with a majority being zero
nearly zero. Given the differences in the patterns and in the magnitude and noise
a properly
levels, the results are considered very good. This demonstrates that
trained neural network has robust ability to recognize patterns.
The false target rate depended on a threshold value used. Since the neural
network output could not always be a perfect binary number with one bit being
1 and the rest being 0, a threshold must be defined. When an output had a
bit
greater than the threshold, the output was considered
to indicate an identified
402
Tan
Table 4 PerformanceoftheNeuralNetworkRecognizer
for Patterns of Different Magnitude and Noise Levels
Data
sequence
Target
rate
(%)
Error target
rate
97-100
99- 100
95- 100
99- 100
0-0.17
0
100
100
IO0
88- 100
0
0
0
0-2. I7
79- 100
0-5.67
(%)
0-0.17
0-0.17
pattern corresponding to that bit. It is not surprising that the threshold value had
10 shows the false target rate
significant effects on the false target rate. Figure
as a function of threshold value for three magnituddnoise combinations. A high
to a high false target rate.
A low threshold value
threshold value would lead
would reduce the false target rate but increase the error target
rate; in other words,
with a lower threshold value, the recognizer would more easily claim data as
recognizable patterns but also more easily misclassify them. Selecting an appropriate threshold value is, therefore, important. From this work, a threshold value
of 0.6 appeared most reasonable.
Threshold value
Fig. 10 False target rate as a function of threshold value (a value above which a target
was considered identified),
New Techniques
403
The performance of the recognizer differed for different data sequences as

shown in Table 4 andFig. 10. Forexample,theperformancemeasuresfor
magnitudehoise combinations y,,l(t),y?.?(t),and yi3(t) were different. The differences were evidently dueto the amount of data used for neural network training.
Three hundred input vectors from R,?(t),100 from y,,,(t), and50 from y3,3(t) were
used to train the neural network. Consequently, the test results for yz,?(t) were
the best with a target rate of 100% and a zero error target rate (Table 4). The
a target rate of 97% andan error target
results for y,,l(t) were slightly worse with
rate of 0.17%. The results for y3.3(t)were the poorest with a target rate of 79%
and an error target rate of 5.67%. The false target rates in Fig. 10 also show the
same trend. This shows, as can be expected, that the performance of the neural
network recognizer for a pattern depends on the amount of training data drawn
from that pattern.
2. ExperimentalDataExample
The methods and procedures described in the last section were usedin an experimental application in food extrusion. Although 15 patterns have been identified
and considered common in the literature, experimentally acquiring data of those
patterns from a specific process is not always practical. First, the responses of
quality variables depend on the process dynamics. It is usually difficult to determine by experimental trial and error what disturbances would result in a certain
pattern in quality data. On the other hand, patterns induced by disturbances frequently occurring in a process may somehow differ from the representative patterns described in the literature. Second, to obtain sufficient data of various pata huge and exhaustive array of experiments
terns for neural network training,
may be necessary, which can be practically unfeasible.
The food extrusion process is expensive to run and cannot be subjected to
extensive experimentation. To facilitate the experiment process and to minimize
the number of experiments necessary, process modeling was used to advantage.
A major source of process variation was identified and perturbed with a pseudorandom binary sequence (PRBS). A quality variable of interest was measured
and modeled as a function of the disturbance variable. The model was used to
determine the typeof disturbance that would rendera certain quality data pattern,
and further experiments were then conducted.The model was also used to generate pattern data for neural network training. Finally, the neural network recognizer was tested on the experimental data.
a. The Process. Afoodextruder is a high-temperature,short-timeprocess that can transform a variety of raw materials into intermediate and finished
In the
productssuch as ready-to-eatfoods, flat breads,andbreakfastcereals.
404
Tan
process, food materials are pressed through a barrel by a set of screws. They are
heated, pressurized, and subjected to shear. The materials often become highpressure, high-temperature, and gelatinized extrudate when they reach the die.
After exiting the die, the extrudate expands and is cut into a desirable size.
The product size is one of the most important quality attributes of many
extruded food products. Since itreflects the degree of expansion and bulk density
of expanded products andis relatively easy to measure, it is a routinely monitored
quality variable for quality control in food extrusion. For constant material feed
rate, cutter speed, and other processing conditions, the product size depends on
the degree of expansion, whichis affected by material properties. Since variations
rein material properties are inevitable, the product size varies. Based on past
search, a major source of process disturbance is feed moisture content. It significantly affects quality attributes such as product size.
h. EquipmentandExperiments.
An APV-BakerMPF-50125twin-screw
food extruder wasused. The extruder consistedof two setsof intermeshing screw
elements of different geometry fitted in an enclosed barrel. A die was mounted
at the end of the barrel. The screws were driven by a 28 kW DC motor, and the
screw speed could vary from 0 to 500 rpm. The feeder was a KTRON model
T35 twin-screw volumetric feeder.
An IVEK Digifeeder system was used
to inject
water into the barrel. The extruder had nine zones with independent barrel temperature controllers. The screw speed, feed rate, moisture addition rate, and cutter
speed were controlled by a host microcomputer.
To measure the extruded product size on-line, a computer vision system
1 color frame
was developed. The system consisted of a Data Translation DT-287
grabber, a DT-2878 advanced processor, two programming libraries (AURORA
and AIPL), a Sony DXC- 15 I CCD color video camera and a Sony PVM-I 342Q
color video monitor hosted by a microcomputer. Each digitized image frame had
512 X 480 pixels. The pixel value range was 0-255 (3 X 8 bit resolution). The
same exposure and focal distance were
used for all images. The selected exposure
was such that the image intensity histograms were roughly centered
at the middle
of the full-scale range (0-255), which gave the best resolution and clearest images. The focal distance was set so that a desired number of product samples
could fit in the image frame. An image processing algorithm was developed in
the C programminglanguagetosegmenttheproductobjectsfromthebackground, identify each sample individually, compute the size (side view area in
of several samples as
square millimeters) of each sample, and take the average
the measured product size.
A sampling mechanism was fabricated to bring a numberof samples from
the product streamto the camera view area
at each sampling instant.The sampling
a
mechanism and the vision computer acted as slaves to the host computer. On
signal from the host, the sampling mechanism would take samples, and the vision
New Techniques
201
405
401
601
801
1001
1201
Time (s)
Fig. 11 The pseudo-random binary sequence (PRBS).
computer would capture a sample image, compute the product size, and transmit
the measurements to the host computer. The sampling period used was 4 s.
The extrusion experiments were performed around an operating condition of screw speed of 300 rpm, feed rate of 45 kg/h, and total moisture content
of 19% (wet basis). Yellow corn meal was used to make a puffed product.
The barrel temperature profile was held constant as described in Chang and Tan
(22).
To produce the effect of natural moisture variations in feed materials, the
moisture addition rate was perturbed with various disturbances while the product
size was measured. First, a pseudo-random binary sequence (PRBS) was designed
and applied to the moisture addition rate to excite the process across its frequency
bandwidth. The PRBS signal is shown in Fig. 1 1 . The amplitude of the disturbance was such that the total extrudate moisture content varied between 17 and
21% (wet basis). The PRBS data were used to develop a model of product size
versus feed moisture content. From the model, other types of disturbances were
selected and applied to the moisture addition rate for more experiments around
the same average operating conditions of the process.
c. Process Modeling. The response of product size to moisture disturbance was modeled with the following ARMAX (auto-regressive movingaverage with auxiliary input) model:
(1
+ a,q- + a2q-*)A(t)= (mlq- + m2q-)M(t - d) + (1 + c,q-)E(t)
( 1 1)
where A(t) is product size (mm), M(t) is moisture content (%, wet basis), E(t)
is a white noise sequence, d is time delay (s), and a,, a?, m,, m!, and c I are constant
coefficients.
The model was developed using the PRBS experiment data. The model
structure and time delay were determined using a systematic search algorithm
406
Tan
Table 5 Model Coefficients and Time Delay
27-0.904
-2.980
13.774
-0.081
-0.863
(22). The recursive least-squares algorithm in Matlab was used to determine the
coefficients. The time delay and coefficients are shown in Table 5.
d. Puttertz Recognizer Training and Testing. From the model (Eq. [ 1 I ] )
it was found that a sinusoidal wave, a square wave, and a step disturbance would,
respectively, induce a cycle, a mixture, and a sudden shift pattern in the product
size data. These three disturbances were then applied to moisture addition rate
in new experiments. The amplitude (or magnitude) of the disturbances was the
same as that used for the PRBS disturbance (17-21 %). The period of the sinusoidal and square wave disturbances was 360 s.
One set of experimental data is plotted in Fig. 12. The data approximately
exhibit the three patterns. There are significant noisesin the data, which resulted
from natural disturbances and were not part
of the patterns resulting from the
special disturbances. After a zero-phase-shift digital filter was used to remove
the noises, the filtered data and those predicted by the model are in good agreeof the special disturbances
ment. This shows that the model described the effects
very well.
Pattern data were generated with Eq. ( I I ) for the same three disturbance
inputs as those used in the experiments. The data were used to train the neural
network pattern recognizer by following the same procedures described in Sec.
1II.D. Data sets measured on-line with the computer vision system were used to
test the trained neural network. The target rate was more than 98% for the cycle
pattern, 95% for the mixture pattern, and 91% for the sudden shift pattern. The
false target rate was zero.
These results further verify the usefulnessof the procedure used to develop
the neural network pattern recognizer. In addition, a process model derived from
multifrequency perturbation is helpful to save experimental efforts.
IV.
REAL-TIMESTATISTICALPROCESSCONTROL
As mentioned in Sec. 111, it is very desirable to have real-time process control

because any time delay in process adjustment can result in a significant amount
of low-grade products or wastes. This is especially true for high-throughput con-
New Techniques
407
..
350
m i
200
181
240
361
541
72 1
901
1081
4
1
271
91
181
541
361
451
Time (s)
Fig. 12 Comparison of model-predicted patterns with measurements: (top) cycle, (middle) mixture, and (bottom) sudden shift. Dotted lines are measured, thicker solid lines are
filtered measurements, and thinner solid lines are predicted.
Tan
408
tinuous processes that cannot be stopped and restarted frequently. Implementation

of real-time SPC would require means for on-line measurement of quality variables, automated detection of abnormalities, and automated determination and
implementation of corrective actions.
Many techniques describedin this book can be used for on-line food quality
measurement. Automatic detection of excessive deviations is simple, and automatic recognition of abnormal patterns is achievable (Sec. 111). With the modern
computer-based process modeling and control techniques, appropriate process
corrective actions can be automatically determined and implemented. Instrumenin improving
tal measurements of food quality therefore have great potential
quality control in food processes.
in Tan et al. (23)
In this section, we use an example application reported
to demonstrate the use of instrumental food quality measurements in the implementation of an automated, real-time statistical process controller.
A.
ProcessandEquipment
The application process was twin-screw food extrusion. The process and equipment used were the same as those described in Sec. 1II.D. Yellow corn meal was
used to make a puffed product, and the quality variable of interest was product
to measure the product side view
size. The computer vision system was used
area in square millimeters, and length and width in millimeters.
B. ProcessVariations
During an extrusion run, feed material properties can vary considerably. For example, the moisture content of the corn meal used in this work could vary from
9 to 13% (wet basis)as a result of differences in batch, supply source, and storage
conditions. This variationcan significantly affect the product size and other quality attributes. To shorten the experimentalrun time, the effects of material moisture variation on product size was demonstrated
by introducing a disturbance into
the moisture addition rate. The disturbance was a sequence of step changes as
shown in Fig. 13, which caused the overall moisture content to vary from
17 to
21%.
The variations in product area are shown in Fig. 13. The dotted line in the
figure shows the area variations when a single sample was measured at every
sampling instant. The plot indicates that the variations consisted of two major
components: a fast or high-frequency variation on top of a slow trend. The fast
component of variationwasduetorandomornaturaldisturbances,andthe
slow component was due to
anassignableorspecialdisturbance,whichwas
moisture change in this case.
Natural disturbances are undeterminable, and thus random variations can-
409
New Techniques
400 r
..""
__
"_
50
Single-sampleoreo
Sinqle-somole
oreo
Subgroup-averagedare0
Sudgroup-averaged
Mosturecontent
Mosture
content
350
300
_.
250
200 I
"""
I"""
30
20
.""_I
1 50
50
1 00
150
200
250
Subgroup or Sample Number
Fig. 13 Product area variations resulting from natural and special disturbances (changes
in moisture content).
not be eliminated. For process control, it is important to detect the existence of

special disturbances. Subgrouping (low-pass filtering) is often used to separate
the effects of the two types of disturbance. Subgroups are chosen so that within
a subgroup variations are considered to be only due to natural disturbances and
A rational
betweentwosubgroupsvariationsareduetospecialdisturbances.
subgroup is chosen in various ways depending on the manufacturing process.
For a continuous process, the key factor for subgrouping is the subgroup size
(number of samples), which determines if the subgroup possesses the properties
described above. Proper selection of the subgroup size usually requires
many
trials. For this work, a subgroup size
of 10 was found appropriate through experiat
ments. When 10 samples were taken to compute a subgroup-averaged area
every sampling instant, the area variations areas shown by the solid line in Fig.
13. The fast or high-frequency variations associatedwith natural disturbances are
largely filtered out by the averaging operation. It is clear from the solid line that
the subgroup-averaged area was inversely related to moisture content.
In addition to identificationof special patterns in quality data, control charts
are used to determine if a process exhibits excessive deviations. TheX Shewhart
control chart is widely employed to monitor the subgroup mean
of a quality
variable by examining if it is between an upper control limit (UCL) and a lower
control limit (LCL) (16). The UCL and LCL are usually expressed as:
UCL = X
+ aR
LCL = X - a R
Tan
. ~ . ..
Length
Width
Moisture c o n t e n t
UCL=23.8
h
h.
Subgroup Number
Fig. 14 Productdimensionvariationsresultingfrommoisturecontentchanges
and LCL are upper and lower control limits, respectively).
(UCL
where X is the grand average of a measured quality variable over a long run
(average of subgroup means), a is a constant depending on the subgroup size,
and R is the average of within-subgroup ranges. Since R indicates the magnitude
by Eqs. (12) and
of random variations in a process, the control limits defined
(13) reflect the process capability to maintain uniformity in a quality variable.
The control limits were determined through experiments. For the product
area, X = 300 mm? and R = 80 mm'. For a subgroup size of 10, a = 0.308
( 16). Then,
UCL = 325 mm2 and LCL = 275 mm'
for the subgroup-averaged product area (solid line
in Fig. 13). Figure 13 is the
of
Shewhart control chart for the product size (area) without implementation
corrective actions. The size was out of the control limits because of the moisture
variation and the process was not in a state of statistical control.
The product length and width variations are shown in Fig. 14. The length
reduced with an increasein moisture content or vice versa. For the product length,
X = 22 mm, R = 6 mm, and a = 0.308 (subgroup size = lo), which give
UCL = 23.8 mm and LCL
= 20.2 mm
As shown in Fig. 14, the product length was also out of its control limits as a
result of the moisture variation.
The product width exhibited little variation relating to moisture change as
New Techniques
41 1
shown by Fig. 14. The size variations were almost exclusively reflected on the
product length. As a result, improved control of the product width was unnecessary for minimizing the effects of moisture variation on size uniformity of the
test product.
C. CorrectiveAction
To bring the process to a state
of statistical control, appropriate corrective actions
must be implemented. Sincethe process was continuous, quick actions are important. Upon detection of a state out of statistical control, corrective actions were
determined by using a simple feedback control scheme. The size measurement
by the vision system was used as the feedback signal to determine a proper cutter
speed to compensate for the effects of moisture variations.
The block diagram of the control system implementedis shown in Fig. 15.
In the figure, product size refers to either product area or length depending
on
which one of the two is of interest for control. The process block stands for a
functional relationship (transfer function) from cutter speed to product size. The
disturbance effect block represents the unknown relationship from moisture conof the controller was to minimize product size
tent to product size. The role
variation under the disturbance of moisture variations.
Since the process dynamics between cutter speed and product size
was
simple, the following PI (proportional and integral) controller was considered
appropriate for determining corrective actions:
u(t) = u(t - 1 )
+ K,[e(t)
e(t
-111 + K,e(t)
(14)
Cmtent
Effect
Cutter
Cmtroller
speed
Process
P r a h t %e
Fig. 15 Block diagram of the vision-based real-time process control system.
412
Tan
350
33
~~~3~s_""""""""_"""
E
E
E
E
v
0
31
300
LCL=275
250
- 25
. . . . . ,*.,. . . . . . . . . . . . . . . . . UCL=_7_338,
,.,.%
, I. \,......_,
*.\ ,l-s.l... ,,, ......_-___
:4, ,.' .
.- ,.,~
'<
.
...:*
\,
a:.;
;%.'e:..
,I
SI
-21
~.I\.
2oo"""""""-
150
"""""_
19z
212
LCL=20.2:
""_
-,
I""
:..lzz-,l
100
1 50
g,
- 17
Molsture content
50
19
"""
- __ Area
.....- .- L e n g t h
""
1 00
23
\"
.._.I.,
200
15
- 13
11
250
Subgroup Number
Fig. 16 Shewhart control chart showing performance of the real-time

LCL are upper and lower control limits, respectively).
SPC (UCL and
where u is the cutter speed control signal, e is the difference (error) between the
K, and K, are gain
desired product size (process mean) and the measured size,
constants, and t stands for the current sampling instant or subgroup number.
The proportional and integral gains, K, and K,, could be designed if a process model were known. As an alternative, they were determined through experimental tuning as widely practicedin industry by using the Ziegler-Nichols tuning
procedure (see, e.g., Ref. 24). The two controller gains were determined as K,=
0.01, K, = 0.03 for product area control andK, = 0.14 and K, = 0.42 for product
length control.
The vision-basedcontrolsystemwasimplementedon-lineandtested
against moisture disturbances. Figure 16 shows a control chart for both product
area and length when the process was subjected to the same moisture disturbance
shown in Figs. 13 and 14. The control chart illustrates the performance of the
system.
In comparison with the uncontrolled curves in Figs. 13 and 14, the system
improved the product size uniformity significantly. The controlled product area
and length were always within UCL and LCL, meaning that the process was in
a state of statistical control in spite of the moisture disturbance. The controlled
curves do not have anidentifiablepattern,indicatingthatthecontrolsystem
largely eliminated the variations caused by the special disturbance.
Figures 17 and 18 are, respectively, the product area and length histograms
with and without the vision-based process control. The histograms show that the
413
New Techniques
0.14
0.12
0.10
0.08
x
W
3
0-
EZd
Uncontrolled .
Controlled
.
0.06 -
LL
0.04 ,0.02
0.00
240 250 260 270 280 290 300 310 320 330 340 350 360 370 380
Area (mm*mm)
Fig. 17 Histograms of product area with and without the real-time SPC.
0.1 4
0.1 2
0.1 0
Uncontrolled
Controlled
6 0.08
3
0-
0.06
0.04
0.02
0.00
1 61 71 8
1 9 20 21
21
22 2 3 24 25 26 27 28 29
Length (mm)
Fig. 18 Histograms of product length with and without the real-time
SW.
Tan
414
uncontrolled product area and length gathered in three clusters corresponding to

the three distinct levels of the moisture disturbance. Within a cluster, the product
size also varied; but the roughly normal shape of the histogram for each cluster
shows that the within-cluster variation was a result of random disturbances. The
standard deviations of uncontrolled product were 27.7 mm' for area and 1.96
mm for length, indicating considerable ranges of variation. The controlled histograms, on the other hand, are essentially limited to the middle clusters despite
the presence of the disturbance. Their roughly normal shapes indicate that the
effects of the special disturbance were mostly eliminated or the process performance was near optimal in terms of product size uniformity.The standard deviations of the controlled product were 7.6 mm' for area and 0.64 mm for length,
which represented, respectively, 73% and 67% reductions over the uncontrolled.
The results show a considerable improvement of product size uniformity by the
application of the vision-based process control system.
V.
FUTURETRENDS
This chapter describes some recent developments in the application

of instrumental measurements for food quality analysis and control. Many
of the concepts
and techniques are new to food applications. Active future research and developments are expected.
Fuzzy set and neural network techniques have great potentialin food quality data analysis. The discussion in this chapter represents only some groundwork
of a fuzzy-set-based paradigm for food quality data analysis and demonstrates
how fuzzy set and neural network techniques may lead to a natural way for food
quality data interpretation. We will see many more future research efforts geared
a fundamentally sound
towardsrefining the methodology and procedures into
and user-friendly system. The method will be tested with various quality datato
establish its reliability and consistency. Computer software applications will appear to facilitate the practical use of the methodology.
Quality data pattern recognition by neural networks will see more and more
applications. Along with the steadily improving availability
of nondestructive
instrumental means for quality measurements, such automated techniques will
play an increasingly important rolein quality control in the food industry. Future
research in this area will emphasize multivariate or multiattribute cases.
Automated real-time statistical process control is the future of SPC. With
the development of computer technology and pattern recognition techniques such
as those discussed in this chapter, automated real-timeSPC has become a reality.
We will see increased integration of process control with SPC andSQC (statistical quality control). In other words, process control will be implemented
in the
New Techniques
415
context of quality control. Furthermore, as food quality is ultimately judged by

the consumers, SPC systems, which are usually based on instrumental quality
measurements, will be increasingly linked with sensory and consumer responses.
Fuzzy set and neural network techniques are important for this link.
REFERENCES
J Tan, X Gao, DE Gerrard. Application of fuzzy sets and neural networksin sensory
analysis. J Sensory Stud 14: I 19-138, 1999.
2. SS Stevens.Mathematics,measurementandpsychophysics.In:
SS Stevens,ed.
Handbook of Experimental Psychology. New York: Wiley, 195 I .
3 . H Stone, JL Sidel. Sensory Evaluation Practices. 2nd ed. San Diego, CA: Academic
I.
Press,1993.
4. GJ Klir, NB Yuan. Fuzzy Sets and Fuzzy Logic: Theory and Applications. Englewood Cliffs, NJ: Prentice Hall, 1995.
5. LA Zadeh. Fuzzy sets. Information Control
8:338-353, 1965.
6. HW Lincklaen Westenberg, S De Jong, DA Van Meel, JFA Quadt, E Backer,RPW
Duin. Fuzzy set theory applied to product classification by a sensory panel. J. Sensory Stud 4155-72, 1989.
7. WM Dong, HC Shah, FS Wong. Fuzzy computations in risk and decision analysis.
Civ Eng Syst 2:201-208, 1985.
8. WM Dong, FS Wong. Fuzzy weighted averages and implementation of the extension
principle. Fuzzy Sets Syst 21:183-199, 1987.
9. CC Lee. Fuzzy logic in control systems: fuzzy logic controller, part 11. IEEE Tran
Syst Man Cyber 20:419-435, 1990.
IO. J Tan, Z Chang. Linearityand a tuning procedure for fuzzy logic controllers. Trans
ASAE37:973-979,1994.
11. B Kosko. Neural Networks and Fuzzy Systems. Englewood Cliffs, NJ: Prentice Hall,
1992.
12. DE Gerrard, X Gao. J Tan. Determining beef marbling and color scores by image
processing, J Food Sci 61:145-148, 1996.
13. AMSA. Guidelines for Meat Color Evaluation. Chicago: American Meat Science
Association, I99 1.
14. CRauwendaal.SPC-StatisticalProcessControl
in Extrusion. NewYork:Hanser
Publishers,1993.
15. GW Sturm,SAMelnyk, MA Yousry,CLFeltz,JEWolter.Sufficientstatistical
process control: measuring quality in real time. In: JB Keats, DC Montgomery, eds.
Statistical Process Control in Manufacturing. New York: Marcel Dekker, 1991.
16. WS Messina. Statistical Quality Control for Manufacturing Managers. New York:
Wiley,1987.
17. JA Swift. Development ofa knowledge-based expert system for control chart pattern
recognition and analysis. PhD dissertation, Oklahoma State University, Stillwater,
OK. 1987.
416
Tan
18. CS Cheng. Group technology and expert systems concepts applied to statistical process control in small-batch manufacturing. PhD dissertation, Arizona State University, Tempe, AZ, 1989.
19. HB Hwarng. Back-propagation pattern recognizers for; control charts: methodology
and performance. Computers Indust Eng 24:219-235, 1993.
Co. StatisticalQualityControlHandbook.NewYork:Western
20.WesternElectric
Electric Co. 1958.
21. J Tan, Y Sun. Quality data pattern recognition for on-line statistical process control.
Proc. 4th Intl. Sym. on Automatic Control of Food and Biological Processes, Goteborg, Sweden,1998.
22. Z Chang, J Tan. Determination of model structure for twin-screw food extrusion 1:
Multi-loop. Trans IChemE 71(C2): 1 1-19, 1993.
23. J Tan, Z Chang, F Hsieh. Implementation of an automated real-time statistical process controller. J Food Proc Eng 19:49-61, 1996.
24. D Seborg, TF Edgar, DA Mellichamp. Process Dynamics and Control. New York:
Wiley,1989.
Index
Absorbance, 6, 17, 119, 220

Absorption, 18, 24, 30, 12 1
coefficient,139
Absorptivity, 6, 8. 348
Acoustic:
image, 223
microscopy, 223
Adenosine triphosphate (ATP), 100
Adenosine monophosphate (AMP),
363
Aflatoxin,124
bruise, 20,21, 110. 143,150, 151
DLEintensity,109
firmness, 244, 249, 252, 270-275
gloss, 20
maturity, 19
spectral characteristics, 18
stiffness, 255
sugar, 19
watercore.21,137,143,147,156
Apricot,DLEintensity,103,109
Artificial intelligence, 40, 74, 92
Attenuated total reflectance (ATR), 14,
15
Avocado firmness, 256, 274. 275

Bacteria:
Catnpylobacter, 129
Escherichiu coli. 361
[Bacteria]
Listeria spp.. 361
Salrnonellu, 129, 36 1
Yersinia enterocolitica, 361
Banana:
DLEintensity,103,106-109
gloss, 20
maturity, I 1 1
Bayesianclassifier,154,155, 158
Beans, color evaluation, 74
Beef
color and marbling, 391-393
fat content, 26
grading, 150
Beer-Lambert Law, 6, 8, 139
Bell pepper:
DLE intensity, 109, I 1 1
gloss, 20
picking, 150
orientation and shape, 1 10
Bioluminescence, 363-364, 370
Biosensors, 335, 337
acoustic, 350
amperometric, 357, 370
applications in food industry, 359
biocatalyst, 338-339
biocatalytic membrane, 356
biocomponent, 338-339, 345-346
biomimetic, 369-370
417
418
[ Biosensors]
biomolecules, 339
biorecognition, 344
electrochemical, 353
optical, 348, 370
potentiometric, 370
Blueberry, firmness, 270
BOD (biological oxygen demand),
337
Boltzmann constant, 203, 205
Bound water, 201
Butter:
composition, 28-29
creep, 306
Camera:
area-array, 55
BCCD, 57-58
calibration, 72
CCD, 55-57, 87, 90, 122, 124
CID, 56
CMOS, 55-57
digital, 58
frame transfer, 56
line-scan, 55, 58, 90
progressive scan, 58
TDI (time delay integrate), 55, 90
Cantaloupe:
firmness, 272
maturity,112
Carotene (carotenoid), 20, 109, 1 I O
Cam-Purcell pulse sequence, 179, 183
Carr-Purcell-Meiboom-Gill(CPMG),
179, 195- 199
Cavitation, 2 19
Cheese:
Cheddar, 307-3 I3
cut time, 227
fat globules, 84-86
meltability, 309
mozzarella, 304
process American, 307
shreds, 66
Cherry:
firmness, 272-273
ripeness, sugar, 19
index
CHESS (chemical shift selective)
sequence,190
Chlorophyll:
content, 99, 107-1 IO, 126, 257
lossldegradation,17, 19, 99,109
Chocolate:
blooming, 43, 56
melting profile, 226
Chromaticity:
coordinates, 4-5, 19
diagram, 5
CIE (Commission de Internationale de
I'Eclairage), 4-5, 19, 69
CLSM (confocal laser scanning microscopy), 83
Cod fillet:
firmness, 248
sealworms and bones, 223
Color:
calibration, 7 1
diagram, 5
index, 22
memory, 1
model. 5
Munsell, 5
rendering index (CRI), 70-71
temperature, 43
Colorimeter, Agtron, 19
Computer tomography (CT), 139- 147, 159
dual energy gamma, 159
Computer vision, 39-92, 1 IO, 130,
404-41 1
Corn:
color evaluation, 74
defects, 23. 24
extrudate characteristics,68
quality factors, 1I O
shape inspection, 66-67
Cracker, shape inspection, 67
Creep compliance, 291 -295, 301 -3 I O
retardation time, 294-308
Cucumber, chilling injury, 1 1 1
Cytometry, 129, 36 I , 362
Deborah number, 290
Debye-Stockes theory, 205
Index
Delayed light enussion (DLE),99- I 16, 130

decay,103, 105, 257
discovery, 99
luminescence, 101
Discrete cosine transform(Dff), 148. 149
Discrete Fourier transform (DFT), 148
DNA probe, 344, 359, 360, 361
DRIFTS (diffuse reflectance infrared
Fourier transform spectroscopy), 15
DSC (differential scanning calorimeter),
210, 227, 320, 324
Dynamic:
mechanical (thermal) analysis (DMA,
MDTA), 320-324
rheometer, 297-299, 3 12, 313
tests, 301, 304
viscosity, 301
Egg:
blood spot, 75
defects, 25
MR image,186
shell color, 25
yolk, 5
Eggplant, gloss, 20
Elastography, 232, 237
Electrochemical sensor, 339
Electrochemiluminescence, 360
Electromagnetic:
radiation, I 17- I 18, 2 I7
spectrum, 1-3, 6, 31, 118
Electronic nose, 365-370
ELISA, 351-352, 362
thermistor ELISA (TELISA). 352-353
Emulsion:
creaming, 23I
flocculation, 23I , 3 16
gel, 316-317
oil-in-water, 3 16
Energy:
acoustic, 221
attenuation, 53
chemical, I O 0
sound, 220
vibrational, 253
Expert systems, 63
419
Fast Fourier transform (FFT), 256, 257

Feature extraction, 41, 61, 62, 65-66
variant and invariant methods, 66
Fiber optics, 31, 44, 52-55, 340
Field effect transistor (FET), 35 1-352
immuno FET (IMFET), 351-352,355
ion-sensitive FET (ISFET), 347, 351
Fixed path length resonator, 222
Fluorescence, 99, 100, 1 16- 130
auto and induced, 1 18- I27
cytometry,129
image, 86, 122
immunosensors, 349-350, 371
labeling,129
microscopy,123,129
Food safety, 3 1, 39, 191, 209, 237, 287,
319, 320, 335, 337, 370
Frame grabber:
calibration, 72
gain and off set control, 57-59
Free induction decay (FID), 175, 194.
196, 212
multiexponential decay, 196
French fry, color, 73, 75
Fresnel equations, 7
Fruit juice, solids concentration, 126,
224
FTIR (Fourier transform infrared), I ,
11-15, 22-31
Fuzzy logic (sets), 63, 74, 75, 92, 154,
156, 379-392
defuzzification, 389
Gel point:
determination, 3 13
gel time, 133, 315
sol-gel transition, 3 14
Winter-Chambon method, 314-315
Genlocking, 61
Glass transition temperature, 201-212,
304, 3 19-323
MRI mapping, 209
Gloss, 20
Grain:
admixture, 23
moisture content, 26
420
Grapefruit, surface defects, 22

Gray level quantization, 144
HAACP (hazard analysis critical control
point), 335, 337-338
Handling:
data, 39. 86
postharvest, 20, 257
Herz contact theory, 266
Hookes law, 289, 295
Hookean, 292-300
HSI (hue, saturation, intensity) system,
42, 68-74, 542
Hysteresis, 29 I , 342
Illumination system, 42-54
Image:
acquisition, 41
blur, 43, 90
features, 65-66, 86, 143
Fourier transform, 182
histogram, 73
morphology, 64-65, 147-148
noise, 143
reconstruction, 84, 144
segmentation, 6 1-62, 73
texture, 67-68
thinning, 65
thresholding, 73
understanding, 41, 63
Injury:
chilling. I I I , 258
impact, 2 I
mechanical. 20, 15I
Interferometer,13.14.222.339
Ion-sclectivc electrode, 355
IR (infrared), 1-2. 12- 19, 27-30. 86
imaging. 86, 88
spectroscopy, 1 1 - 12, 26, 28
Isochronal. 295-296
Karhunen-Loeve transform, 148- 149
Kclvin-Voight model. 294-295.307-308
Kiwifruit:
tirmness, 48, 249-259, 270-275
maturity, 188
Index
[Kiwifruit]
modulus, 266
MR image, 188
texture, 256
Knowledge base, 41, 62-63
Kubelka-Munk, 8, I O
Larmor equation, 166
frequency,168,171, 180
Laser air-puff firmness detector, 244,
259, 265
Laser Doppler vibrometer, 256-257,
275-276
Lemon:
color grade, I I O
DLE intensity, 107-1 10
Ligase chain reaction (LCR), 360
Lighting:
arrangements, 43
sources, 43
strobe, 43, 90
Linescan,139-141, 153
Loss tangent (tan 6). 300, 3 14-315,
321 -322
Lubein, 20
Lycopene,109
Machine vision, 39-92
Magness-Taylor (MT) puncture test,
244-246, 253. 263, 269, 271
Magnetic dipoles, 166
Mango, firmness, 256, 274
Maxwell model, 292, 294. 301, 306
Melon. firmness. 244. 248, 254, 270,
274-275
Microorganismlmicrobial:
activity. 201, 209

contamination. 337, 362, 364, 370
infestation,I23
in-line analysis. 337
rapid analysis, I29
testing, 361 -362
toxins,123-124,335.348
Microscopy:
acoustic. 223
confocal laser scanning (CLSM). 83
index
421
[Microscopy]
electron, 83
fluorescence,123,129
3-D, 83
Microwave, 275
Milk:
coagulation, 84, 310. 31 1
gel, 310-312
composition, 26, 30, 224
Modulus:
bulk, 219
complex, 300
elastic, 219, 264
loss (viscous), 298
shear, 289
storage, 298, 320, 322
Young's, 232, 259, 289
Moisture content, measurement,
26
Molecular imprinting, 368-370
MRI (magnetic resonance imaging),
165,179-185, 2.57
Munsell color atlas. 5
Muscle food, fat thickness, 232
Muskmelon, DLE intensity, 113
Mycotoxins, I23
Oil:
adulteration, 28
composition,28-29,127,129
Olives, DLE intensity, 109
Onion:
diseases,137, 151
DLE intensity, 109
firmness, 273
gloss, 20
line scan image, 153
separation from clods, 252
On-line:
control, 63
firmness sorting, 278
inspection,142
moving scene analysis, 89
quality and safety monitoring, 371
statistical process control, 396
viscometer, 236
Optical density (OD), 7. 18, 20-26, 1 19
Orange:
DLE intensity, 102-1 10
firmness, 247, 252, 270, 273
gloss, 20
juice, 22
surface defects. 22
Nectarine:
DLE intensity, 109
firmness, 252
Neural network, 63, 74-75 92, 154,
157, 379-398
Neuro-fuzzy systems, 75
Newton's law, 289
Newtonian. 292-293, 297
NIR (near-infrared), I , 9, 1 1 , 18, 21 31, 54, 57, 86, 88. 115, 227, 244.
253, 280
image, 86-88
instruments, I2
sensors. 160
spectroscopy. 26, 257, 275
NMR (nuclear magnetic resonance),
165-21 1
imaging.165
spectroscopy,165,166,288
Papaya:
DLEintensity,103,108
maturity,112.113
Partial least squares (PLS), 379
Pattern recognition, 63, 64, 154, 232,
396-406
Pea, firmness, 275. 276
Peach:
DLE intensity,102, 109
firmness, 246, 249, 259. 270-275
ripeness, 19, 1 13
Peanut:
maturity, 20
moisture content, 26
Pear, firmness, 249, 270-273
Penetrometer, 245
Persimmon,DLEintensity.103.105,
109,113
Phase locked loop (PLL), 60
422
Phosphorescence, 7, 100
Photoluminescence, 1 13, 1 16
photoluminography,I30
Photonintensity,139
Photosynthesis, 100, 130
Pineapple, surface color, 22
Pistachio, color classification,75
Pixel, 41
square, 55, 59
jittering, 60
Plancks constant, 1 I , 167
Plum, DLE intensity, 109, 113
Poissons ratio, 251, 264, 266
Polymerase chain reaction, 360
Pomegranate, DLE intensity,109
Pork, muscle quality, 24-25
Potato:
color evaluation, 74
diseases, 23
firmness, 273
hollow heart, 23, 234
Principal component analysis (PCA),
22, 379
discriminant analysis, 22
Proportional and integral controller,
41 1
Prune:
ripeness, I9
sorting, I 14
Quanta, 1 I8
quantumnumber,166,168
Quenching,119,120
Quartz resonator sensor, 366
Radiofrequency (RF) pulse, 167- 174,
183,185,190
Reduced mass, 1 1-12
Reflection:
body, 9
coefficient, 223
diffuse, 7, 10, 17
regular, 7
specular, 7, IO, 43
total internal, 52, 348-349
Refractive index, 4, IO- 12,52-53.348
Index
Relaxation time, 17 I - 172, 184- 185,
189, 220, 289-290, 305
Retardation time, 294, 305
RGB (red-green-blue) system, 5, 42,
68-74, 393
Rice:
degree of milling, 24
gelatinization, 24
quality factors, I 10
RNA probe, 359-360
SAOS (small amplitude oscillatory
shear), 288, 296, 300-319

temperature and frequency sweep,
309
Scattering:
classical theory, 220
coefficient, 8, 12
losses, 227
of muscle, 25
Raleigh, I I9
sound, 229
thermal, 229
viscous, 229
Semiconductor metal oxide chemoresistive sensor, 366
Signal-to-noise ratio, 22, 343
Sing-around method, 22 1
Snells law, 6, 52
Soybeans, quality factors, 1 10
Spectrophotometer,12,26,113,120
Spectroscopy:
ATR.14,15
diffuse reflectance, 9, 15
dynamic mechanical, 320
ITIR, I , 13, 14, 15, 25
NIR, I , 12, 22, 25, 88, 257
optical, 220
photo correlation, 348, 350
ultrasonic. 217, 227, 236
visible, 1
Spin-lattice and spin-spin relaxation,
17 1 - 172, 192-209, 257
spin-echo pulse, 177- 178
Spirits, alcohol concentration, 224
Starch gelatinization, 24
Index
Statistical process control (SPC), 379,

396,406, 414
statistical quality control, 414
Strawberry:
firmness, 270
maturity, 188
MR image, 188
Stress relaxation, 293, 305, 306
Surface plasmon resonance (SPR), 348,
370
Tea leaves, DLE intensity, 108, I 1 I
Texture, 233-244
profile analysis (TPA), 287, 288
3-D measurement techniques:
food quality analysis, 39
stereo, 8 1-82
structured light, 79-81
time of flight, 76-78
triangulation, 78-79
Time constant, 171, 172, 204-209,
257
Time-temperature superposition, 304
Tomato:
DLE intensity, 103- I 10
firmness, 247, 270, 275
gloss, 20
internal color, 19
maturity,I87
stern and blossom end, 1 I O
Transducer types, 338-347
piezoelectric. 250, 253, 272-274,
338, 348, 350
Transmittance, 6, 7, 10
Ultrasonic:
attenuation, 219, 222, 227, 230, 233
firmness sensing, 254
image, 23 1
423
[Ultrasonic]
pulse-echo method, 22 I , 232
pitch-and-catch method, 22, 222
through-transmission method, 22 I
velocity, 219,221,224,226,227,230
UV (ultraviolet), I , 54, 86, 100, I 15117,121,124
Viscoelasticity, 288-289, 301, 309. 324
Viscosity, 205. 243, 289, 307, 350
measurement, 3 12
complex, 3 18
Voxel,140,142
Water activity, 191,192
Waves:
acoustic, 291
evanescent, 348, 350
longitudinal, 2 18
mechanical, 2 17-2 18
shear, 2 18, 236
sound, 217
surface acoustic (SAW), 236, 350
Wheat:
fat content, I28
protein content, 28
smut content, 24
Xanthophyll, 20
X-ray:
absorption coefficient, 141,146
dualenergy,160
image, 86, 89, 139,141,150, 158
photons,144
source,139,146
Yogurt, gel stiffness, 3 I I

Nondestructive Food Evaluation Techniques To Anyaluze Properties and Quality Food Science and Technology

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Nondestructive Food Evaluation Techniques To Anyaluze Properties and Quality Food Science and Technology

Enviado por

Direitos autorais:

Formatos disponíveis

Nondestructive

FOOD SCIENCE AND TECHNOLOGY

Owen R. Fennema University of Wisconsin-Madison

1. Flavor Research: Principles and Techniques, R. Teranishi, 1. Hornstein,

70.Handbookof Fruit ScienceandTechnology:Production, COmpOSitiOn,

97. Seafood Enzymes: Utilization and Influence on Postharvest Seafood

Additional Volumes in Preparation

This book is printed on acid-free paper.

Marcel Dekker, Inc.

Copyright 0 2001 by Marcel Dekker, Inc. All Rights Reserved.

In the food industry, we are inherently limited

4) to cover the wide span

Optical Methods: Visible, NIR, and FI'IR Spectroscopy

Delayed Light Emission and Fluorescence

X-Ray Imaging for Classifying Food Products Based on Internal

5 . Nuclear Magnetic Resonance Techniques and Their Application

Joseph Irudayaraj Department of Agricultural andBiologicalEngineering,

Optical Methods: Visible, NIR, and

Gunasekaran and lrudayaraj

to a broad band of the electromagnetic spectrum and in establishing indices of

II. LIGHT ANDCOLORMEASUREMENT

Fig. 1 The electromagnetic spectrum.

If the spectral distribution throughout the visible region is unequal, then

lrudayaraj Gunasekaran and

It is obvious from the equations above thatx y + z = 1. The tristimulus values

C. MunsellSystem and Atlas

Gunasekaran and lrudayaraj

When light falls on an object, it may be reflected, transmitted, or absorbed (Fig.

According to its transmittance properties, an object may be transparent,

The attenuation of the transmitted rayin a homogeneous, nondiffusing, absorbing

where a is called the absorptivity. [If c is expressed in mol/L and b in cm, a is

energy such as fluorescence, phosphorescence, delayed light emission, heat, etc.

(n2/nl)?cos 8, - [(nz/nl)- sin? 8J1/?

The regular reflectance R = Ri

and for normal incidence (8 =

where R, = absolutereflectance of an infinitelythicklayer,k

where R and R = reflectance of the standard and the sample

versely proportional to particle size) and moisture content. Equation (14)

Gunasekaran and lrudayaraj

of a solid in the presence of a nonabEvaporation of volatile solutions

NIRAND FTlR SPECTROSCOPY

IR spectroscopy has been used as an analytical technique for almost a century

where 2) = frequency, h = Planck's constant, k = force constant, p = reduced

Gunasekaran and lrudayaraj

C -0 absorbs at a different frequency thanC=O. The C =O bond has a higher

Table 1 CharacteristicFunctionalGroups ofFoodComponents

Fig. 3 Schematicdiagram of theMichelsoninterferometer.

Gunasekaran and lrudayaraj

radiation is altered in a complex pattern as a functionof mirror movement. Thus,

Fig. 4 Illustration of attenuated total reflectance. 8, = angle of incidence. 8,

where h = wavelength of the radiation in the internal reflectance element, 0 =

Gunasekaran and lrudayaraj

Fig. 5 Schematic of DRIFTS with diffuse reflectance accessory.

by introducing Fourier transform methods (sensitivity and good signal-to-noise

Single wavelength measurement-the optical property

Gunasekaran and lrudayaraj

To identify diseased potatoes, Muir et al. (39) used

where the subscripts A, B, C, and D represent wavelengths at which

The CIE tristimulus values X, Y, and Z or the chromaticity coordinates x,

Gunasekaran and lrudayaraj

6. Detection of External and Internal Defects

7. Fruits and Vegetable Products

Gunasekaran and lrudayaraj