Você está na página 1de 6

Performance Evaluation of Different Texture Models

for Texture Classification


G. Madasamy Raja

Dr.V.Sadasivam

Associate Professor, Department of IT,


PET Engineering College, Vallioor,
Tirunelveli District, Tamilnadu 627 117, India.
anushpriya2004@yahoo.com

Principal, PSN College of Engineering & Technology,


Tirunelveli, Tirunelveli District.
Tamilnadu-627152, India.

Abstract Texture classification is one of the significant as well


as useful applications in image processing area. Texture
classification is applied in various pattern recognition tasks. In
the literature, so many texture models have been presented but
the problem remains challenging in feature vector generation and
comparison of similarity. This paper evaluates the performance
of some conventional texture measures namely Texture Spectrum
and Local Binary Patterns which have been successfully used in
various applications and of some new promising approaches
proposed recently namely Local Ternary Patterns and Optimized
Local Ternary Patterns. All these texture models are tested by
applying texture classification methods on different image sets
from the standard Brodatz texture database. Experimental
results are then analyzed in order to evaluate texture
classification performance of the above mentioned texture models
based upon the classification accuracy and execution time.
Index Terms Kullback leibler distance, Texture spectrum
(TS), Local Binary Patterns (LBP), Local Ternary Patterns
(LTP), Texture Classification, Optimized Local Ternary Patterns
(OLTP).

I. INTRODUCTION
The analysis of texture in images is one of the ways for the
recognition of objects. Texture classification is one approach
among them for object recognition or pattern recognition. This
paper evaluates the texture classification performance of some
texture measures which have been successfully used in various
applications and of some new promising approaches proposed
recently using two popular benchmarks namely classification
accuracy and execution time. Texture classification evaluation
is not a new area of work, however previous work has either
compared too few algorithms or used very small number of test
images that makes it difficult to come to a conclusion about the
accuracy of results [1]. Texture methods can broadly be
classified into statistical, geometrical, structural, model-based
and signal processing features [2]. A research report about
many texture methods and their performances has been
presented by VanGool et al. [3]. Reed and Buf [4] have
presented a detailed survey of the various texture methods used
in image analysis. Randen and Husoy [5] have also done
research about these kinds of surveys and expressed their

conclusion. Co-occurrence texture model was declared the best


by the research conducted by Ohanian and Dubes [6] as well
as by Conners and Harlow [7]. Ojala et al. [8] compared a
range of texture methods and declared that the best
performance was achieved by the grey level difference method.
II. TEXTURE MEASURES USED IN THIS STUDY
A. Texture Spectrum
The Texture Spectrum method was first introduced by He
and Wang [9] and subsequently extended its usage for many
applications [10]. The texture spectrum method is based on the
computation of the relative intensity relations between the
pixels in a small neighborhood and not on their absolute
intensity values. The texture spectrum which characterizes the
original image with the help of images texture characteristics
will be the result after the application of the texture spectrum
methodology to a given digital image. Texture Spectrum
method uses a basic concept called Texture Unit. A Texture unit
is characterized by eight pixels each of which has one of three
possible values (0,1,2) obtained from a neighborhood of 3*3
pixels. If the intensity value of the central pixel is considered as
X0 and the intensity value of each neighboring pixel as X i, the
set that is considered as the smallest complete unit of the under
consideration image is:
TU = {X0, X1, X2, , X8} where

0 if Xi X 0

Ei 1 if Xi X 0 for i 1, 2, . .8. (1)


E = The i element of texture unit set TU = {E , E ,E }.
2 if X i X 0

From the above formula, each element can be assigned one of


i

th

three possible values so the total number of possible texture


units for the eight elements can be estimated as 38 = 6561. In
the texture spectrum method, the eight elements may be
ordered differently. If the eight elements are ordered clockwise

as shown in Fig. 1, the first element may take eight possible


positions from the top left (a) to the middle left (h) and then the
6561 texture units can be labeled by the following formula
under eight different ordering ways (from a to h).
8

N TU Ei * 3i 1

(2)

i 1

Fig. 2. Transforming a neighborhood to a texture unit with the


texture unit number.

Where NTU is the texture unit number. Fig. 2 gives an


example of transforming a neighborhood to a texture unit with
the texture unit number.
B. Local Binary Pattern
As the Texture spectrum method uses very large number of
patterns, another texture model namely Local Binary Patterns
(LBP) operator was introduced by Ojala et al. [11] which was
again based on a 3 3 neighborhood. It operated with eight
neighboring pixels using the center as a threshold. The final
LBP code was then produced by multiplying the threshold
values by weights given by powers of two and concatenating
the results binomially that results in a pattern. This process is
shown in Fig. 3.
In order to describe texture structure in a larger scale, the
original LBP operator is extended to vary the neighborhoods.
The gray-scale invariance local image texture with P neighbors
and R radius is characterized by

P-1
LBPP,R = u(ti tc)2i
i=0

Fig. 3. An example of LBP computation

C. Local Ternary Patterns (LTP)


As LBP may be sensitive to noise especially in uniform
image regions a 3-valued pattern instead of a binary pattern
namely Local Ternary Patterns (LTP) was introduced by Tan
and Triggs [12]. When a 3x3 neighborhood around a centre
pixel in an image is considered, the LTP operator takes the
form

(3)

LTP 3n s(i n i c ),
n 1

where tc is the intensity of the center pixel, ti is the intensity of


the neighborhood i, and u(x) is the step function. P represents
angular resolution and R represents spatial resolution in the
above equation (3). The LBP P,R operator produces 2P different
patterns. So, for a 3x3 neighborhood a total of 256 (2 8)
different pattern strings can be generated. An entire image can
be represented by a pattern histogram of 256 elements, which
can be used as the texture descriptor.

a b c
h
d
g f

Fig. 1. Eight clockwise, successive ordering ways of the 8 elements of the


texture unit.

and

- 1 if u i c t ,

s(u) 0 if i c t u i c t ,
1 if u i t ,
c

(4)

where t is a user-defined threshold, n is the number of


neighboring pixels surrounding the center pixel c, i c is the pixel
value of c and in are the pixel values of n. In LTP, the difference
between the centre pixel and neighboring pixel is encoded by
three values (1,0,-1) according to a predefined and fixed
threshold t, and the length of the pattern histogram is very high
(38). To reduce the dimension of the pattern histogram, the
ternary pattern is converted into binary pattern, by splitting it
into its positive and negative parts. Two separate histograms
for both positive and negative components are calculated and
then the results are concatenated.

D. Optimized Local Ternary Patterns (OLTP)


In OLTP, a texture image can be decomposed into a set of
small units called patterns. As the texture model OLTP uses
only optimal patterns (selected number of patterns), the
length of the pattern histogram is maintained as optimum
compared to other texture models. In OLTP, a pattern is
represented by eight elements, each of which has one of
three possible values (0,1,5) that are obtained from a
neighborhood of 3x3 pixels. Let ic, i1, i2, ,i8 be the pixel
values of a local 3x3 neighborhood region, where ic is the
value of the central pixel and i1, i2, ,i8 are the pixel values
of its 8 neighbors. The following equation (5) defines the
process for converting a local 3x3 neighborhood into its
pattern representation (P),

0 if i n (1 )i c

P(i c , i n ) 1 if (1 )i c i n (1 )i c (5)
5 if i (1 )i
n
c

where is a small scaling factor and it is assigned as 0.05.


Figure 4 shows an example of transforming a sample 3x3 local
neighborhood into a pattern unit by using equation (5). For a
3x3 local neighborhood region, the total number of different
pattern strings will be 6561(38). In other words, Eq. 5 will
deliver any one pattern string from a set of 6561 different
pattern strings for a 3x3 local neighborhood region when
OLTP texture model is used. So, by using this OLTP texture
model a complete texture image can be described by a pattern
histogram of 6561 bins that represents the occurrence
frequency of pattern strings over the texture image.
Further, after many reliable experiments, it is observed that for a
3x3 local neighborhood of a texture image, only few pattern
strings among these 6561 different pattern strings, are
frequently occurring patterns and all other pattern strings are not
so. So it is meaningless to allot separate bins for all the pattern
strings because it may end in wastage of memory and wastage
of time. Further, the analysis of these frequently occurring
pattern strings found that all these pattern strings have uniform
circular structure and they are rotation invariant. To identify
these frequently occurring pattern strings which are rotation
invariant, let us consider, a uniformity measure U which
corresponds to the number of spatial transitions circularly
among the sub patterns in the pattern strings. The uniformity
measure U is defined as

25 17
168
2
2
10005510
15 25
108
6
6
24 38
332
4
0

n 2

1 if | A B | 0
where F(A,B) =
0 otherwise

0
0

(a)

(b)

(c)

For example, the pattern string 11111111 has U value of 0,


the pattern string 00000055 has U value of 2 and the pattern
string 55501111 has U value of 3. Patterns with U value of
less than or equal to three are considered as Uniform
Pattern Strings and totally 45 uniform pattern strings are
available. The implementation of a new idea to calculate the
Transition Length () among the uniform patterns and the
introduction of a new concept called Level of Optimality
(Lopt) have paved the way to select the subset of uniform
patterns which are designated as optimal patterns. The
elements of the uniform pattern strings are either 0 or 1 or 5
or any combinations of these values.
If the number of occurrences (cardinality) of 0, 1 and 5 are
represented by Card(0), Card(1) and Card(5) respectively,
then the level of optimality for a particular uniform pattern
string can be computed as,

L opt

1 min Card(0), Card(1), Card(5)


max

(7)

where refers to the transition length among the sub


patterns in the uniform pattern strings. Table 1 contains the
transition lengths of various sub patterns that are available
in the uniform pattern strings.

Optimized Local Ternary Patterns (OLTP) texture model


that uses only optimal set of uniform patterns to represent a
local image texture is defined as:

P(i , i

OLTP

i 1

25

(6)

Fig. 4. Example Of Transforming A 3x3 Neighborhood to A Pattern


String in OLTP Texture Model.
(a) 3x3 Local Region (b) Pattern Matrix (c) Pattern String

U F(P(i c , i1 ), P(i c , i 8 )) F(P(i c , i n ), P(i c , i n 1 ))

) if U 3 and L opt 2
otherwise

(8)
This recently proposed texture model, OLTP [13] uses only 24

unique optimal patterns for texture representation and it groups


all other patterns under one label 25 which are termed as suboptimal patterns. Therefore the dimension of pattern spectrum
has been reduced from 6561 to 25, that too with optimal set of
patterns. The occurrence frequency of all the optimal patterns
over a whole texture image is termed as OLTP Pattern
Spectrum or OLTP Pattern Histogram, which reflects the
texture information of the corresponding image.
III. TEXTURE CLASSIFICATION
In order to evaluate the performance of the texture models
in the process of texture classification, nine texture images
have been selected from the image set of Brodatzs natural
images [14]. These images were selected because they are
broadly similar to one another and also that they resemble parts
of remotely sensed images. All the selected images are shown
in Fig. 5 and all of them consist of 256 x 256 pixels. A
quantitative study was carried out using a supervised
classification over the selected nine texture images of Fig. 5. A
sample sub images of 30 x 30 pixels was selected within each
texture. Using a window of 30 x 30 pixels, together with a step
of two pixels in the row and column, the full image was
processed and each central pixel of the window was assigned
to one of the four texture classes. Here, the histogram of nine
images using Texture spectrum, LBP, LTP and Optimized
Local Ternary Patterns texture models were calculated within a
window of 30 x 30 pixels. The kullbak-leibler distance was
used as the similarity measure and the difference between two
pattern spectrums was considered as the distance between
texture classes. The supervised classification results of the
images with time taken for execution are shown in section IV.
TABLE I.

Fig. 5: The Brodatz Texture Images Used In the Classification


Experiments
a) pressed calf leather b) brick wall c) straw matting
d) loose burlap e) woven aluminium wire f) lizard skin g) reptile skin
h) herringbone weave i ) raffia

IV. RESULTS AND DISCUSSION


Table 2 shows the classification accuracy of the
experiments for the various texture methods.

DETAILS OF THE TRANSITION LENGTH () FOR THE SUB


PATTERNS OF UNIFORM PATTERN STRINGS
TABLE 2: TEXTURE CLASSIFICATION ACCURACY RESULTS

S.N
o

Transition Length
Sub Pattern
Transition Length ()
type

01

10

15

51

05

50

015

510

105

10

150

11

051

12

501

Image
Name
pressed calf
leather
brick wall
straw
matting
loose burlap
woven
aluminium
wire
lizard skin
reptile skin
herringbone
weave
raffia

TS

LBP

LTP

OLTP

92.41

89.32

94.98

98.96

91.60

86.13

93.48

96.66

82.45

81.12

91.90

92.56

90.54

89.45

94.45

95.54

92.49

89.30

94.12

96.96

80.33
87.65

78.92
85.11

84.88
89.78

88.33
90.87

93.02

89.08

94.09

98.60

79.09

73.33

86.55

88.23

As seen from table 2, when classification accuracy is


considered, it is concluded that the best performance was
achieved by the texture model OLTP followed by LTP. The
performance of LBP was the poorest. The texture method TS
did not perform well compared to LTP and OLTP. In the case
of LTP texture model, 91.58 percent average classification
accuracy is achieved where as in the case of OLTP 94.08
percent average classification accuracy was received which is
the maximum in these experiments. The valid reason for this
achievement is LTP uses fixed threshold for finding the
patterns in the texture where as OLTP uses flexible threshold to
find the pattern in the texture. Among the four texture models
under consideration, the recently proposed OLTP texture model
delivers superior performance with the average classification
accuracy of 94.08 percent.
Figure 6 illustrates the average computational cost for the
selected four texture methods. As TS model is using very large
number of bins in the pattern histogram (6561), it took more
execution time for giving the result. In the case of LBP, since
this texture model uses 256 patterns it also took more time than
LTP and OLTP for giving the output. Both LTP and OLTP gave
the result in more or less same time. But by taking into account
the classification accuracy with execution time, it can be
observed that OLTP outperforms all other texture models in the
process of texture classification because it gave best
performance in both classification accuracy and execution
time.

V. CONCLUSION
In all the discussed experiments, OLTP with 24 unique
optimal set of patterns, gave the superior performance as
indicated by results. From the results it is found that recently
proposed texture model OLTP was showing the best
performance in both benchmark factors that is classification
accuracy and time complexity. Regarding further work, this
comparative study can be extended for the process of texture
segmentation also. More over this study can be further
continued by including more texture measures as well as more
texture images. In future, this comparative analysis may follow
the same procedure with different distance measurement or
different classification algorithms for both texture image
classification as well as textures segmentation. This study uses
the texture images as it is but these images can be rotated or
degraded by any noise signals for further study to assess the
quality of various texture models.
REFERENCES
[1]
[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]
[10]
[11]

[12]

[13]
Fig. 6: Computational cost (in seconds) for four texture models
for the Classification Experiments

[14]

G.Smith and I.Burns, Measuring texture classification


algorithms, Pattern Recognition Letters 18:1495 1501, 1997.
M.Tuceyran and A.K. Jain, Texture analaysis, in Handbook of
pattern Recognition and Computer Vision, C.H. Chen, L.F. Pau
and P.S.P Wang (Eds.), chapter 2, 235 276, World Scientific,
Singapore, 1993.
L.VanGool, P. Dewaele and A. Oosterlinck, Texture analysis,
computer vision, Graphics and Image Processing,29:336357,1985.
T.R.Reed and J.M.H. Buf, A review of recent texture
segmentation and feature extraction techniques, Computer
Vision, Image Processing and Graphics, 57(3):359-372, 1993.
T. Randen and J.H. Husoy, Filtering for texture classification:
A comparative study, IEEE Transactions on Pattern Analysis
and Machine Intelligence, 21(4):291-310, 1999.
P.P. Ohanian and R.C. Dubes, Performance evaluation for four
class of texture features, Pattern Recognition, 25(8):819-833,
1992.
R.W. Conners and C.A.Harlow, A theoretical comparison of
texture algorithms, IEEE Transactions on Pattern Analysis and
Machine Intelligence, 2(3):204 222, 1980.
T. Ojala, M. Pietikainen and D. Harwood, A comparative study
of texture measures with classification based on feature
distributions, Pattern Recognition, 29(1):51-59, 1996.
L. Wang and DC. He, Texture Segmentation Using Texture
Spectrum, Pattern Recognition, Vol. 23, pp. 905-910, 1990.
DC. He and L. Wang., Texture Features Based on Texture
Spectrum, Pattern Recognition, Vol. 24, pp. 391-399, 1991.
T. Ojala, M. Pietikainen and T. Maenpaa, Multiresolution grayscale and rotation invariant texture segmentation with local
binary patterns, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 24, no. 7, pp. 971987, 2002.
X. Tan and B. Triggs, Enhanced Local Texture Feature Sets for
Face Recognition under Difficult Lighting Conditions., IEEE
Transactions on Image Processing, 19(6), pp. 1635-1650, 2010.
G. Madasamy Raja and V. Sadasivam, Optimized Local
Ternary Patterns: A New Texture Model with Set of Optimal
Patterns for Texture Analysis, Journal of Computer Science,
vol. 9, no. 1, pp. 114, 2013.
P. Brodatz, Textures, A Photographic Album for Artists and
Designers, Dover Publications, New York, 1966.

Você também pode gostar