JOURNAL OF COMPUTING, VOLUME 3, ISSUE 4, APRIL 2011, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 101

Fingerprint Recognition Using

Extended Fuzzy Hypersphere Neural Network
M. H. Kondekar, Dr. U. V. Kulkarni & S. S. Chowhan

Abstract— Presently personal identification using fingerprint is one of the most reliable and popular biometric recognition
method. An accurate and consistent classification algorithm can significantly reduce fingerprint matching time. This paper
describes Extended Fuzzy Hypersphere Neural Network (EFHSNN) with its learning algorithm, which is an extension of Fuzzy
Hypersphere Neural Network (FHSNN). The EFHSNN uses Manhattan distance instead of Euclidian distance. The experimental
results for PolyU HRF fingerprint database show that EFHSNN is superior and yields 100% recognition rate along with less
training and recall time.

Index Terms— Biometrics, Fuzzy Neural Network, Hypersphere, Fingerprint Recognition, FingerCode.

——————————  ——————————

1 INTRODUCTION

T he increased focus on confidentiality and safety of

information in databases, personal identification has
become very important topic in today’s network so-
ciety. Biometric indicators have an advantage over tradi-
tional security identification methods, because these inhe-
rent attributes cannot be easily stolen. Up to the present,
there are many biometric features that are used for people
identification, like iris, face, retina, voice, gait, palm print
and fingerprint. The convenience of current electronic
applications has led to an explosive increase in their use.
E-banking, electronic fund transfer, online shopping and
virtual auctions are just some applications prevalently
used by the public. Trust, as a result, has become more of
an issue [1].
Fingerprints are widely used as personal identification
technique around the world [2] due to its distinguished
features from others. Fingerprints are fully formed at
about seven months of fetus development and finger
ridge configurations do not change throughout the life of
an individual except due to accidents such as bruises and
cuts on the fingertips [3].This property makes fingerprints
a very attractive biometric identifier. Fingerprint features
are permanent and fingerprints of an individual are
unique.
The fingerprint recognition algorithms can be broadly
classified into minutiae-based and FilterBank-based algo-
rithms. The minutiae-based matching algorithms first
extract the local minutiae such as ridge endings and ridge
bifurcations from the thinning image [4] or the gray scale
image, and then match their relative placement in a given
fingerprint with the stored template. A number of match-
ing techniques are available in the literature including
point-based matching [4], graph-based matching [5], and
————————————————
 M. H.Kondekar is with the College of Computer Science and Information
Technology, Latur, 413512, Maharashtra, India.
 Dr. U. V. Kulkarni is with the S.G.G.S.I.E.T.,Nanded, 431606, Maharash- process is based on FilterBank-based FingerCode feature
tra, India.
 S. S. Chowhan is with the College of Computer Science and Information
Technology, Latur, 413512, Maharashtra, India.
string-based matching. Although the minutiae-based
matching is widely used in fingerprint verification, but it
has problems in efficiently matching two fingerprint im-
ages containing different number of unregistered minu-
tiae points. Further, it does not utilize a significant por-
tion of the rich discriminatory information available in
the fingerprints. The FilterBank-based algorithm [6, 7, 8]
uses a bank of Gabor filters to capture both local and
global information in a fingerprint as a compact fixed-
length FingerCode, which is suitable for matching and
storage. Thus, it overcomes some of the problems with
the minutiae-based matching algorithms. So, here we
have used FilterBank-based algorithm [8] for efficient and
correct fingerprint feature extraction.
In this paper, we have applied EFHSNN classifier
which is an extension of Fuzzy Hypersphere Neural
Network (FHSNN) proposed by Kulkarni et al [9, 12] to
the problem of fingerprint recognition based on Finger-
Code feature data. The EFHSNN utilizes fuzzy sets as
pattern classes in which each fuzzy set is a union of fuzzy
set hyperspheres. The fuzzy set hypersphere is a n-
dimensional hypersphere defined by a center point and
radius with its membership function. The rest of the pa-
per is structured as follows. The feature extraction me-
thod in our work is introduced in Section 2. Then Sections
3 gives a brief introduction for the architecture of the
EFHSNN, followed by its learning algorithm in section 4.
Section 5 examines the experimental results of the clas-
sifiers on fingerprint data set. Conclusions are made in
Section 6.
2 FILTERBANK BASED FINGERPRINT FEATURE
EXTRACTION
In this paper fingerprint feature extraction is done by us-
ing PolyU HRF Fingerprint database images of 320*240
sizes at 1,200 dpi resolution. The feature extraction
extraction algorithm by Jain et al [8] which consists of
following stages.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 4, APRIL 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 102

2.1 Reference Point Location 3 EXTENDED FUZZY HYPERSPHERE NEURAL

Fingerprints have many visible landmark structures NETWORK
and a combination of them could be used for establishing
The EFHSNN consists of four layers as shown in Fig. 1(a).
a reference point. Jain, Prabhakar, Hong, and Pankanti
had defined the reference point of a fingerprint as the The first, second, third and fourth layer is denoted
point of maximum curvature of the concave ridges in the as FR , FM , FN and FO respectively. The FR layer accepts
fingerprint image. The location of reference point is main- an input pattern and consists of n processing elements,
ly dependent on good quality of image, for graceful han- one for each dimension of the pattern.
dling of local noise in a poor quality fingerprint image; The FM layer consists of q processing nodes that are
the detection should necessarily consider a large neigh- constructed during training and each node represents
borhood in the fingerprint image. For locating a reference
hypersphere fuzzy set characterized by hypersphere
point of a fingerprint Local ridge orientation is usually
specified for a block rather than at every pixel; an image membership function [12]. The processing performed by
is divided into a set of non overlapping blocks and a sin- each node of FM layer is shown in Fig. 1(b).The weight
gle orientation is defined for each block. between FR and FM layer represents centre points of
the hyperspheres.
2.2 Filtering and FingerCode Feature Extraction As shown in Fig. 1(b), C j  (c j1 , c j 2 ,....., c jn )
Fingerprints have local parallel ridges and valleys, and  represents center point of the hypersphere m j . In addi-
well  defined  local  frequency  and  orientation.  Properly  tion to this each hypersphere takes one more input de-
tuned Gabor filters [10, 11] can remove noise, preserve the  noted as threshold T , which is set to one and the weight
true ridge and valley structures, and provide information  assigned to this link is j .j represents radius of the
hypersphere m j , which is updated during training. The
contained in a particular orientation in the image. Before 
center points and radii of the hyperspheres are stored in
filtering the fingerprint image, it is normalized to the re‐ matrix C and vector  , respectively. The maximum size
gion  of  interest  in  each  sector  separately  to  a  constant  of hypersphere is bounded by a user defined value  ,
mean  and  variance.  Normalization  is  performed  to  re‐ where 0    1 is called as growth parameter that is
move the effects of sensor noise and gray level distortion  used for controlling maximum size of the hypersphere
due to finger pressure differences.  and it puts maximum limit on the radius of the hyper-
sphere.
An even symmetric Gabor filter has the following gen‐
Assuming the training set defined
eral form in the spatial domain: as R  {Rh | h = 1,2, …, P} , where
  1  x'2 y '2   Rh  ( rh1, r h 2 ,...., rhn )  I n is the hth pattern, the mem-
G ( x, y; f , )  exp  2  2   cos(2fx' ) (1)  bership function of the hypersphere node m j is
2 
  x '  y'  
              
x'  x sin  ycos                             (2)  m j ( Rh , C j ,  j )  1  f (l ,  j ,  )
 
               (4) 
y'  x cos  y sin                          (3)    
Where  f is the frequency of the sinusoidal plane wave 
along the direction   from the  x ‐axis, and   x ' and  y ' are 
the  space  constants  of  the  Gaussian  envelope  along  x'  
and  y '  axes, respectively. Jain et al [9] had performed the 
filtering  in  the  spatial  domain  with  a  mask  size  of  33    
33. In this algorithm Jain, Prabhakar, Hong, and Pankanti 
had  used  eight  different  values  for   
( 0 0 , 22.5 0 , 45 0 , 67.5 0 , 90 0 , 122.5 0 , 135 0 , and 157.5 0 ) 
with  respect  to  the  x ‐axis.  The  normalized  region  of  in‐
terest  in  a  fingerprint  image  is  convolved  with  each  of 
these  eight  filters  to  produce  a  set  of  eight  filtered  fea‐
tures.  These  eight  directional‐sensitive  filters  capture 
most of the global ridge directionality information as well 
as  the  local  ridge  characteristics  present  in  a  fingerprint.  Fig. 1(a). Extended Fuzzy Hypersphere Neural Network.
The  mean  of  each  sector  in  each  of  the  eight  filtered  fea‐
tures  defines  the  components  of  FingerCode  feature  vec‐
tor. The gray level in a sector in a disk represents the fea‐
ture  value  for  that  sector  in  the  corresponding  filtered 
image.
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 4, APRIL 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG
where f ( ) is three‐parameter ramp threshold function  n k  max m j u jk for k  1,2,..........., p         (8)
defined 
Fig. 1 (b). Implementation Extended Fuzzy Hypersphere
Neural Network
  1  x ' 2 y ' 2  
G ( x, y; f ,  )  exp   2   cos( 2fx ' )  (5)
2 
2 
  x '  y'   
and the argument l is defined as,
 n 
  c ji  rhi 
 i 1 
The membership function returns m j  1 , if the input
pattern Rh is contained by the hypersphere. The parame-
ter , 0    1 ,is a sensitivity parameter, which governs
how fast the membership value decreases outside the
hypersphere when the distance between Rh and C j in-
creases. The sample plot of membership function with
centre point [0.5 0.5] and radius equal to 0.3 is shown in
Fig. 2.
Fig. 2. Plot of Modified Fuzzy Hypersphere Membership
Function for = 1
Each node of FN and FO layer represents a class. The
FN layer gives fuzzy decision and output of k th FN node
represents the degree to which the input pattern belongs
to the class n k . The weights assigned to the connections
between FM and FN layers are binary values that are
stored in matrix U and updated during learning as
1 if m j is a hypersphere of class n k 
u jk  
For k =1,0 2,otherwise
. . . . . . ., p and j =1, 2, . . . . . . ., q .              j new   c ji  rhi 
where  m j is the  jth FM node and  n k is the  k th FN node.  Step  3:  If  the  pattern  Rh is i not
Each  FN node  performs  the  union  of  fuzzy  values  re‐
turned by the fuzzy set hyperspheres of same class, which 
is described by Eq. (5). 
q 103
 
j 1
Each  FO  node  delivers  non‐fuzzy  output,  which  is  de‐
scribed by Eq. (6).
0 if m j  T
ok   for k  1,2,..........., p           (9) 
  1if n k  T
where  T  max ( n k ) for k  1,2,....., p . 
4 EFSHNN LEARNING ALGORITHM
The EFHSNN consists of four layers as shown in Fig. 1(a). 
The  first,  second,  third  and  fourth  layer  is  denoted 
as FR , FM ,  FN and  FO respectively.  The FR layer  accepts 
an  input  pattern  and  consists  of  n   processing  elements, 
one for each dimension of the pattern. 
The  supervised  FHSNN  learning  algorithm  for  creating 
fuzzy hyperspheres in hyperspace consists of three steps 
 
1. Creation of hyperspheres, 
2. Overlap test, and 
3. Removing overlap. 
(6) These three steps are described below in detail. 
 
4.1 Creation of Hyperspheres
Given the hth training pair ( Rh , d n ) find all the hyper-
spheres belonging to the class d n . These hyperspheres
are arranged in ascending order according to the dis-
tances between the input pattern and the center point
of the hyperspheres. After this following steps are car-
ried sequentially for possible inclusion of input pat-
tern Rh .
Step 1: Determine whether the pattern Rh is con-
tained by any one of the hyperspheres. This can be veri-
fied by using fuzzy hypersphere membership function
defined in Eq. (4). If Rh is contained by any of the hyper-
sphere then it is included, therefore in the training
process all the remaining steps are skipped and training is
continued with the next training pair.
Step  2:  If  the  pattern  Rh falls  outside  the  hypersphere, 
then the hypersphere is expanded to include the pattern if 
the  expansion  criterion  is  satisfied.  For  the  hypersphere 
m j to include  Rh the following constraint must be met. 
   n 

 c ji  rhi          
 i 1 
     (10) 
Here  we  have  proposed  a  new  approach  for  testing 
expansion of new hyperspheres based on Manhattan dis‐
tance  which  is  sum  of  the  absolute  differences  yields  su‐
perior results as compared to Euclidian distance used by 
FHSNN. 
If the expansion criterion is met then pattern  Rh  is in‐
 cluded as   n 

(7) 
         (11) 
 
1            
included  by  any  of  the 
above  steps  then  new  hypersphere  is  created  for  that 
class, which is described as 
 
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 4, APRIL 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 104

Cnew  Rh and  new 0                   (12) 

 
 
4.2 Overlap Test
The  learning  algorithm  allows  overlap  of  hyperspheres 
from  the  same  class  and  eliminates  the  overlap  between 
hyperspheres from different classes. Therefore, it is neces‐
sary  to  eliminate  overlap  between  the  hyperspheres  that 
represent  different  classes.  Overlap  test  is  performed  as 
soon as the hypersphere is expanded by step 2 or created 
in step 3. 
(a) Overlap test for step 2: Let the hypersphere  mu is 
expanded to include the input pattern  Rh and expansion 
has created overlap with the hypersphere  mv , which be‐ Fig. 3(b). Status of the hyperspheres after removing an overlap in
longs  to  other  class.  Suppose,  step 2.
C u  [ x1 , x 2 ,........., x n ] and  u   represents  center  point 
and  radius  of  the  expanded  hypersphere  and   
 
If the step 3 creates overlap then it is removed by
modifying of other class. Let C p  [ x1 , x 2 ,........., x n ] and
 p, represents centre point and radius of the created
hypersphere and C q  [ x '1 , x ' 2 ,........., x ' n ] and  q, are
center point and radius of the hypersphere of other class.
Then overlap is removed as 
 

Fig. 3(a). Status of the hyperspheres before removing an

overlap in step 2.

C v  [ x'1 , x' 2 ,........., x' n ] and  v ,  are  centre  point  and 

radius  of  the  hypersphere  of  other  class  as  depicted  in 
Fig. 3(a). Then if 
 n 
  cui  rvi   u   v                      (13) 
Fig. 4(a). Status of the hyperspheres before removing an overlap in
step 3.
 i 1 
means those hyperspheres from separate classes are over‐
lapping. 
(b) Overlap test for step 3: If the created hypersphere 
falls inside the hypersphere of other class means there is 
an  overlap.  Suppose  m p represents  created  hypersphere 
to  include  the  input  pattern Rh ,  and  m q represents  the 
hypersphere  of  other  class  as  shown  in  Fig.  4(a).  The 
presence of overlap in this case can be verified using the 
membership  function  defined  in  Eq.  (4).  If  Fig. 4(b). Status of the hyperspheres after removing an over-
m p ( Rh , C p ,  p ) mq ( Rh Cq  q )  1 means  two  hyper‐ lap in step 3.
spheres from different classes are overlapping.   n

 q new    c pi  rqi    (15)
4.3 Removing Overlap  i 1 
If step 2 has created overlap of hyperspheres from sepa-
rate classes then overlap is removed by restoring the ra- 5 EXPERIMENTAL RESULTS
dius of just expanded hypersphere. Let, mu be the ex-
panded hypersphere then it is contracted as
The EFHSNN is implemented using MATLAB 7.0. The
u  u
new new
(14) results are obtained and compared with FHSNN. For
and new hypersphere is created for the input pattern as comparisons we have used fingerprint database. The pro-
described by Eq. (12). This situation is shown in Fig. 3(b). file of the database is as shown in section 2. The timing
analysis of training and recall are depicted in Table 1 and
recognition rates in terms of number of hyperspheres,
 JOURNAL OF COMPUTING, VOLUME 3, ISSUE 4, APRIL 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 105

 
TABLE 1
4. A.K.  Jain,  L.  Hong,  S.  Pankanti,  and  R.  Bolle,  “An  identity  au‐
TIMING ANALYSIS
thentication  system  using  fingerprints”,  Proceedings  of  the  IEEE, 
Classifier Average Training Average Vol. 85, No. 9, pp. 1365‐1388, 1997. 
Time (Seconds) Recall  
Time 5. A. K. Hrechak and I. A. Mchugh, “Automated fingerprint recogni‐
(Seconds) tion using structural matching”, Pattern Recognition, vol. 23, no. 
FHSNN using Euc- 0.463710 4.000131 8, pp. 893‐904, 1990. 
lidian distance  
6. A. K. Jain, S. Prabhakar, and L. Hong, “A multichannel approach 
EFHSNN using 0.450811 3.686820
to  Fingerprint  Classification”,  IEEE  Transactions  on  Pattern 
Manhattan distance
Analysis  and  Machine  Intelligence,  vol.  21,  no.  4,  pp.  348‐359, 
  1999. 
TABLE 2  
PERCENTAGE RECOGNITION RATE WITH FHSNN AND EFHSNN 7. A. K. Jain, S. Prabhakar, L. Hong, and S, Pankanti, “FingerCode: 
Classifier Radius Hyper- Recognition a FilterBank for fingerprint representation and matching”, Proceed‐
spheres Rate ings of IEEE Computer Society Conference on Computer Vision 
and Pattern Recognition(CVPR), vol. 2, pp. 187‐ 193, 1999. 
FHSNN 0.0249 280 100 %  
EFHSNN 0.09 262 100 % 8. A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “FilterBank‐
  based  fingerprint  matching”,  IEEE  Transactions  on  Image 
radius and recognition rate are depicted in Table2. Processing, vol. 9, no. 5, pp. 846‐859, 2000. 
 
9. U.V.  Kulkarni  and  T.R.  Sontakke,  “Fuzzy  Hypersphere  Neural 
As  shown  in  Table  1  the  training  and  testing  time 
Network Classifier”, Proceedings of 10th International IEEE Con‐
(0.450811,  3.686820  seconds  respectively)  of  EFHSNN  ference on Fuzzy Systems, University of Melbourne, Australia, 
classifier  is  fast  compared  to  any  other  classifier.  Also  as  December 2001, pp. 1559‐1562. 
shown in Table 2 the number of hyperspheres (262) is also   
less  in  comparison  with  FHSNN.  Hence,  the  EFHSNN  10. J. G. Daugman, “High confidence recognition of persons by a test of 
statistical  independence”,  IEEE  Transactions  on  Pattern  Analysis 
classifier  gives  better  recognition  rates  in  comparison 
and Machine Intelligance, vol. 15, no. 11, pp. 1148–1161, 1993. 
with  FHSNN  in  terms  of  less  number  of  hyperspheres,   
training and recall time.  11. J. G. Daugman, “Uncertainty relation for resolution in space, spatial 
frequency,  and  orientation  optimized  by  two‐dimensional  visual  cor‐
tical filters”, J.Opt. Soc. Amer. A. vol. 2, pp. 1160–1169, 1985. 
6 CONCLUSIONS  
For effective classification of fingerprint in network socie- 12. B.  B.  M.  Krishna  Kanth,  U.  V.  Kulkarni  and  B.  G.  V.  Giridhar, 
ty, we have demonstrated effectiveness of EFHSNN clas- “Gene  Expression  Based  Acute  Leukemia  Cancer  Classification:  a 
sifier on fingerprint dataset. Experimental results shows Neuro‐Fuzzy  Approach”  International  Journal  of  Biometrics  and 
that the EFHSNN classifier is the most effective compared Bioinformatics, (IJBB), Volume (4): Issue (4) pp. 136–146, 2010.
with FHSNN classifier in terms of less number of hyper-
M. H. Kondekar received the M.C.A. degree from Government En-
spheres, training and recall time. The training time differ- gineering College, Aurangabad in the year 2000. He received the
ence of 0.012899 seconds and testing time difference of M.Phil. Degree in Computer Science from Y.C.M.O. University, Na-
0.313311 seconds is less as compared with FHSNN clas- shik in the year 2009. He is currently working as Lecturer in the Col-
sifier. Thus EFHSNN can be used for huge realistic data- lege of Computer Science and Information Technology, Latur, Maha-
rastra. His current research interests include various aspects of
base recognition purposes where less amount of training Neural Networks and Fuzzy Logic, Pattern Recogntion and Biome-
and testing time is the prime request. Our future work trics.
will be based on exploring clustering techniques using
fuzzy neural approach. Dr.U.V.Kulkarni received the PhD degree in Electronics and Com-
puter Science Engineering from S.R.T.M. University, Nanded in the
year 2003. He is currently working as Head of Computer Science
REFERENCES Engineering Department and Dean of academics, SGGS Institute of
Engineering and Technology, Nanded, Maharastra.
1. P.Meenen, and R.Adhami, “Fingerprinting for security”, Proceed‐
ings of the IEEE Potentials, pp. 33‐38, 2001. 
  S. S. Chowhan received the M.Sc.(CS) degree from Dr. BAM Uni-
2. T.  Song,  Liang  Huang,  Che‐Wei  Liu,  Jui‐Peng  Lin,  Chien‐Ying  versity, Aurangabad in the year 2000. He received the M.Phil. De-
Li,  and  Ting‐Yi  Kuo,  “A  Novel  Scheme  for  Fingerprint  Identifica‐ gree in Computer Science from Y.C.M.O. University, Nashik in the
tion”,  Proceedings  of  the  Second  Canadian  Conference  on  year 2008. He is currently working as lecturer in the College of Com-
Computer and Robot Vision (CRV’05) 0‐7695‐2319‐6/05.  puter Science and Information Technology, Latur, Maharastra. His
  current research interests include various aspects of Neural Net-
3. W.  J.  Babler,  ʺEmbryologic  Development  of  Epidermal  Ridges  and  works and Fuzzy Logic, Pattern Recogntion and Biometrics.
Their  Configurationʺ,  Birth  Defects  Original  Article  Series,  vol. 
27, no. 2, 1991.