PIFS Code Base for Biometric Palmprint Verification by ijcsis


More Info
									                                                  (IJCSIS) International Journal of Computer Science and Information Security,
                                                  Vol. 9, No. 2, February 2011

                                            I Ketut Gede Darma Putra
                              Departement of Electrical Engineering, Faculty of Engineering
                                 Udayana University, Bukit Jimbaran, Bali - Indonesia
                                             email : duglaire@yahoo.com

Abstract — This paper proposes a new technique to extract            resolution images can be used, low cost capture devices
the palmprint features based on some fractal codes. The              can be used, it is very difficult or impossible to fake
palmprint features representation is formed based on position        palmprints, and their characteristics are stable and unique
of range blocks and direction between the position of range          [18].
and domain blocks of fractal codes. Each palmprint
                                                                         Recently, many verification/identification technologies
representation is divided into a set n blocks and the mean
value of each block are used to form the feature vector. The         using palmprint biometrics have been developed
normalized correlation metrics are used to measure the               [2],[3],[4],[5],[11],[12],[13],[18],[21]. Zhang et al. [21]
degree of similarity of two feature vectors of palmprint             applied 2-D Gabor filter to obtain the texture features of
images. We collected 1050 palmprint images, 5 samples from           palmprints. Pang at al. [13] used the pseudo-orthogonal
each of 210 persons. Experiment results show that our                moments to extract the features of palmprint. LI et al. [12]
proposed method can achieve an acceptable accuracy rate              transformed the palmprint from spatial to frequency
with FRR = 1.754, and FAR= 0.699.                                    domain using Fourier transform and then computed ring
                                                                     and sector energy features. Connie at al.[2] extracted the
Keyword; biometrics, fractal codes, fractal dimension,               texture feature of palmprint using PCA and ICA. Wu et
feature extraction, palmprint recognition                            al.[18] extracted line feature vectors (LFV) using the
                                                                     magnitudes and orientations of the gradient of the points
                                                                     on palm-lines. Kumar et al.[11] combined the palmprints
                   I. INTRODUCTION                                   and hand geometries for verification system. Each
    The personal verification becomes an important and               palmprint was divided into overlapping blocks and the
highly demanded technique for security access systems in             standard deviation value of each block was used to form
this information area. Traditional automatic personal                the feature vector.
recognition can be divided into two categories: token-                   In this paper, we propose a new technique to extract the
based, such as a physical key, an ID card, and a passport,           features of palmprint based on fractal codes. This
and knowledge-based, such as a password and a PIN.                   technique is different with the method in [4] and [5].
However these approaches have some limitations. In the
token-based approach, the “token” can be easily stolen or
lost. In the knowledge-based approach, the “knowledge”                               II. IMAGE ACQUISITION
can be guessed or forgotten [21]. In order to reduce the                 All of palm images are captured using Sony DSC P72
security problem caused by traditional methods, biometric            digital camera with resolution of 640 x 480 pixels. Each
verification techniques have been intensively studied and            persons was requested to put his/her left hand palm down
developed to improve reliability of personal verification.           on with a black background. There are some pegs on the
Biometric-based approach use human physiological or                  board to control the hand oriented, translation, and
behavioral features to identify a person. The most widely            stretching. A sample of the hand and pegs position on the
used biometric features are of the fingerprints and the most         black board is shown on Figure 1 (a).
reliable are of the irises. However, it is very difficult to
extract small minutiae features from unclear fingerprints
and the iris input devices are very expensive [19]. Other
                                                                              III. PALMPRINT EXTRACTION AND
biometric features such as of face, voice, hand geometries,
and handwritten are less accurate. Faces and voices can be
mimicked easily, hand geometries and handwritten can be                 This paper used new technique to extract the ROI
faked easily.                                                        (region of interest) of palmprint. This technique consists of
    Palmprint is the relatively new in physiological                 two steps in center of mass (centroid) method. These steps
biometrics [18]. There are many unique features in a                 can be explained as follow.
palmprint image that can be used for personal recognition.           a. The gray level hand image is thresholded to obtain the
Principal lines, wrinkles, ridges, minutiae points, singular              binary hand image. The threshold value was computed
points and texture are regarded as useful features for                    automatically using the Otsu method. To avoid the
palmprint representations [21]. A palmprint has several                   white pixels (not pixel object) outside of the hand
advantages compared to other available features: low-                     object is used median filter.

                                                                47                            http://sites.google.com/site/ijcsis/
                                                                                              ISSN 1947-5500
                                                           (IJCSIS) International Journal of Computer Science and Information Security,
                                                           Vol. 9, No. 2, February 2011

b.   Each of the acquired hand images needs to be aligned
     in a preferred direction so as to capture the same
     features for matching. The moment orientation method
     is applied to the binary image to estimate the
     orientation of the hand. In the method, the angle of
     rotation ( θ ) is the difference between normal axis and
     major axis of ellipse that can be computed as follows.

          1               2 µ1,1                                                            (a)                  (b)                (c)
     θ = tan −1                       
          2            µ 2,0   − µ0,2 

     µ p ,q = ∑∑ (m − m ) (n − n )q
                m      n                                                                (d)              (e)               (f)              (g)

     where    µ p,q                           th
                       represent the (p,q) moment central, and               Figure 1. Extraction of palmprint, (a) original image, (b)
     ( m, n ) represents center of area is defined as                           binary image of (a), (c) object bounded, (d) and (e)
                                                                            position of the first centroid mass in segmented binary and
                1            1
          m=      ∑∑ m , n = N ∑∑ n ,
                N m n
                                                                 (3)         gray level image, respectively, (f) and (g) position of the
                               m n                                           second centroid mass in segmented binary and gray level
     where N represents number of pixel object.                                                  image, respectively.
     Furthermore, the grayscale and the binary image are
     rotated about ( θ ) degree.
c.   Bounding box operation is applied to the rotated
     binary image to get the smallest rectangle which                                         IV. FEATURES EXTRACTION
     contains the binary hand image. The original hand
     image, binarized image, and the bounded image                               There are three main steps to extract the palmprint
     shown in Figure 1 (a), (b), and (c), respectively.                     features based on fractal codes proposed in this paper.
d.   The centroid of bounded image is computed using                        These steps can be explained as follows.
     equation (3) and based on this centroid, the bounded
     binary and original images are segmented with 200 x                    A. Extraction of fractal codes of palmprint images
     200 pixels. The segmented image and its centroid                            Fractal codes of palmprint images are obtained using
     position are shown in Figure 1 (d) and (e).                            the partitioned iterated function system (PIFS) method. In
e.   The centroid of the segmented binary image is                          PIFS method, each image is partitioned into its range
     computed and based on this centroid the ROI of                         blocks and domain blocks. The size of the domain blocks
     grayscale palmprint image can be cropped with size                     is usually larger than the size of the range blocks. The
     128 x 128 pixels. The first and the second positions of                relation between a pair of range block (Ri) and domain
     centroid in binary and gray level image are shown in                   block (Di) is noted as
     Figure 1 (f) and (g).
                                                                                         Ri = wi (Di )                                                (6)
    This method is so simple. This method has been tested
for 1050 palmprint images acquired from 210 persons, and                    wi is contracted mapping that describes the similarity
the results show this method is reliable.                                   relation between Ri and Di, and is usually defined as an
     Before the feature extraction phase, the extracted ROI                 affine transformation as below:
are normalized using normalization method in [11] to                                         xi   a i          bi     0   xi   ei 
reduce the possible imperfections in the image due to non-
uniform illumination. The method is as below:                                            wi  y i  =  ci
                                                                                                               di     0   yi  +  f i 
                                                                                                                                                 (7)
                                                                                             zi   0
                                                                                                               0      s i   z i  oi 
                                                                                                                                
                         φ d + λ      if I ( x, y ) > φ
          I ' ( x, y ) =                                        (4)
                         φ d − λ          otherwise                        where xi and yi represent top-left coordinate of the Ri , and
                                                                            zi is the brightness value of its block. Matrix elements ai,
                                                                            bi, ci, and di, are the parameters of spatial rotations and
                      ρ d {I ( x, y ) − φ}2                                 flips of Di, si is the contrast scaling and oi is the luminance
          λ=                                                     (5)
                                ρ                                           offset. Vector elements ei and fi are offset value of space.
                                                                            In this paper, we used the size of domain region twice the
where I and I’ represents original grayscale palmprint                      range size, so the values of ai, bi, ci, and di are 0.5. The
image and the normalized image respectively, φ and ρ                        actual fractal code pi below is usually used in practice[19].

                                                                                   ((               )(            )                     )
represents mean and variance of the original image
respectively, while φd and ρd are the desired values for                     f i = x Di , y Di , x Ri , y Ri , sizei , θ i , s i , oi                 (8)
mean and variance respectively. This research use φd = 180
and ρd = 180 for all experiments.

                                                                       48                                      http://sites.google.com/site/ijcsis/
                                                                                                               ISSN 1947-5500
                                                                                        (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                        Vol. 9, No. 2, February 2011

where     (xR , y R ) and (xD , y D ) represent top-left
                    i             i                        i          i

coordinate position of the range block and domain block,
respectively, and size is the size of range block. The fractal
codes of a palmprint image is denoted as follow:
              F = U fi                                                                           (9)                                       (a)                            (b)
                         i =1
where N represents the number of the fractal code. The
inequality expression below is used to indicate whether the
range and the relevant domain block are similar or not.

              d ( R, D ) ≤ ε ,                                                                  (10)
                                                                                                                                (c)                 (d)
where d(R,D) represents rmse value, and є is the threshold
                                                                                                             Figure 2. Palmprint feature extraction, (a) original image,
(tolerance) value. The range and the relevant domain block
                                                                                                             (b) Image I, (c) Image I’, (d) block feature representation
is similar if d(R,D) is less or equal than є. Otherwise, the
block is regarded not similar.
                                                                                                            The Figure 2 (d) show the palmprint feature representation
                                                                                                            in 16 x 16 sub blocks. Figure 3 shows example of three
B. Palmprint features representation
                                                                                                            groups of palmprints from the same palm and palms with
    The first step of this method is the forming of angle
                                                                                                            similar/different line structures. The features of these
image A as follows.
                                                                                                            palmprints are plotted in figure 4. The results show that the
     A( j , k ) = α i , j = 1,2,3, K M 1 , k = 1,2,3, K M 2                                     (11)        features of three palm images from the same person are
                                                                                                            close to each other than the features of three palm images
                          yD − yR                                                          ,
     α i = arctan                                   if     j=x            and k = y                         from the different persons with similar or different line
                          xD − xR          i
                                                                 Ri                   Ri
   otherwise, α i = 0                              (12)
where x D , y D
                   represent top-left coordinate of the
                          i                                                                                          V. PALMPRINT FEATURE MATCHING
domain block (see formula (8)) and di represent the angle
between range and domain block. The angle image is not                                                           The degree of similarity between two palmprint
binary image representation. The criterion below are added                                                  features is computed as follows:
to compute the direction α i .                                                                                      d rs = 1 −
                                                                                                                                      (xr − xr )(x s − x s )T (15)
if   xR   < xD          and       yR   ≥       yD        then   αi        = αi                                                   [(x   r   − x r )( x r − x r )
                                                                                                                                                                  ] [(x
                                                                                                                                                                          s   − x s )( x s − x s )

if   xR   > xD          and       yR   ≥       yD        then   αi        = 180 − α i                       where    x r , x s are the mean of palmprint feature xr and xs ,
if   xR   > xD          and       yR   ≤       yD        then   αi        = 180 + α i                       respectively. The above equation computes one minus
                                                                                                            normalized correlation between palmprint feature vector xr
if   xR   < xD          and       yR   ≤       yD        then   αi        = 360 − α i
                                                                                                            and xs. The values of drs are between 0 – 2. The d rs will
if   xR   = xD          and       yR   ≥       yD        then   αi        = 90
                                                                                                            be close to 0 if xr and xs obtained from two image of the
if   xR   = xD          and       yR   ≤       yD        then   αi        = 270                (13)
                                                                                                            same palmprint. Otherwise, the d rs will be far from 0.
                                                                                                                Figure 4 shows comparison of feature component of
The criterion            sizei = min(size) means the palmprint                                              those palmprint shown in figure 3, and their score are listed
features representation is formed practically using the                                                     in Table 1. The matching score of group A are close to 0,
coordinate of the smallest size range block. Later, the                                                     and the matching score of group B and C are far from 0.
representation is filtered as follow.                                                                       The average score of group A, B, and C are 0.1762,
     I ' ( x , y ) = I ( x , y ) ∗ h ( x , y )m x n ,                                          (14)         0.5057, and 0.6452, respectively. It is easy to distinguish
                                                                                                            group A from group B and C using these scores.
h(x,y) is filter which all of its component are one. Figure
2(b) show the palmprint features image of Figure 2(a).

C. Palmprint feature vector
    Palmprint feature vector (V) is obtained by dividing
the palmprint image into 16 x 16 blocks, and for each
block its mean value is computed, so obtained the feature
vector V = (v1 , v 2 K , v N ) , where N = 256,and vi is
                                                                                                                        (a1)                             (a2)                             (a3)
                                                                                                                        Group A: palmprints from the same person
mean value of block i.

                                                                                                       49                                            http://sites.google.com/site/ijcsis/
                                                                                                                                                     ISSN 1947-5500
                                                      (IJCSIS) International Journal of Computer Science and Information Security,
                                                      Vol. 9, No. 2, February 2011

             (b1)             (b2)                (b3)
  Group B: palmprints from different person with similar line


          (c1)                (c2)                (c3)
  Group 3: palmprints from different person with different line

      Figure 3. Example of three groups of palmprint

Table 1 Matching Score of groups A, B, and C in figure 3

                      a1       a2         a3      Average
        a1             0     0.1957     0.1404
        a2          0.1957      0       0.1925     0,1762                                                        (b)
        a3          0.1404   0.1925        0
                      b1       b2         b3      Average
        b1             0     0.5352     0.3056
        b2          0.5352      0       0.6763     0,5057
        b3          0.3056   0.6763        0
                      c1       c2         c3      Average
        c1             0     0.6900     0.6177
        c2          0.6900      0       0.6280     0,6452
        c3          0.6177   0.6280        0

             VI. EXPERIMENTS AND RESULTS                                   Figure 4. Comparison of feature component of the
     We collected palm image from 210 persons from both                 palmprint group shown in figure 2. (a),(b),(c) are feature
sexes and different ages, 5 samples from each person, so               component of group A, B, and C, respectively. Red, green,
our database contains 1050 images. The resolution of hand              blue color are the first, second, and third palmprint in each
image is 640 x 480 pixels. The palmprint images, of size                                    group, respectively.
128 x 128 pixels, were automatically extracted from hand
image as described in the Section 3. The averages of the
first three images from each user were used for training
and the rest were used for testing.
     The performances of the verification system are                               400

obtained by matching each of testing palmprint images                              300
with all of the training palmprint images in the database. A
matching is noted as a correct matching if the two                           v26   200

palmprint images are from the same palm and as incorrect                           100

if otherwise.
                                                                                         300                                                        250
                                                                                               200                                      150
                                                                                         v24         100
                                                                                                             0   0

                                                                              Figure 5. Distribution of three feature components
                                                                                 of 1050 palmprints in feature space

                                                                  50                                       http://sites.google.com/site/ijcsis/
                                                                                                           ISSN 1947-5500
                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                 Vol. 9, No. 2, February 2011

                                                                   method for palmprint verification. The experiment results
                                                                   show that the proposed method can achieve an acceptable
                                                                   accuracy rate with FRR = 1.7544, and FAR= 06998. In the
                                                                   future, we will combine the proposed method with wavelet
                                                                   transformation to extract the feature of palmprint to retain
                                                                   the block operation.

                                                                   [1] Chih-Lung Lin., “Biometric Verification Using
                                                                        Palmprints and Vein-patterns of Palm-dorsum”,
                                                                   [2] Connie T., Andrew Teoh, Michael Goh, David Ngo,
                                                                        2003, “Palmprint Recognition with PCA and ICA”,
                                                                   [3] C.L. Lin, Biometric Verification Using         Palmprints
                                                                        and      Vein-patterns     of    Palm-dorsum,      2004,
                                                                   [4] Darma Putra, IKG., Adhi Susanto, A. Harjoko & TS.
                                                                        Widodo, Palmprint Verification based on Fractal
                                                                        Codes and Fractal Dimensions, Proceedings of the
                                                                        Eighth IEASTED International Conference Signal and
                                                                        Image Processing, Honolulu, Hawai, 2006, 323–328.
                                                                   [5] Darma Putra, Adhi Susanto, Agus Harjoko, Thomas
                                                                        Sri Widodo, 2006, Biometrics Palmprint Verification
               (a)                         (b)                          Using Fractal Method, EECCIS proceedings, Part 2,
                                                                        pp.22-23, Brawijaya University, Malang, Indonesia.
 Figure 6. Performance of verification system,(a) genuine
                                                                   [6] Duta N., Jain A.K., Mardia K.V.,2002, Matching of
and imposter distribution, (b) FAR/FRR/EER with various
                                                                        Palmprints, Pattern Recognition Letters, 23, pp. 477-
                                                                   [7] Ekinci Murat, Vasif V., Nabiyev, Yusuf Ozturk, 2003,
     Table 2. FRR/FAR with various threshold value                      A Biometric Personal Verification Using Palmprint
                                                                        Structural Features and Classifications, IJCI
        Threshold         FRR              FAR                          Proceedings of Intl, XII, Vol.1, No.1.
          0.4386         2.0734           0.4734                   [8] Jain A.K., 1995, Fundamentals of Digital Image
                                                                        Processing, Second Printing, Prentice-Hall, Inc.
          0.4586         1.9139           0.5158
                                                                   [9] Jain A.K., Ross A., and Pankanti S., 1999, A Prototype
          0.4626         1.7544           0.6998                        Hand       Geometry-based        Verification    System,
          0.4746         1.4354           0.9160                        www.research.ibm.com/ecvg/publications.html
          0.4786         1.2759           1.3552                   [10] Jain A.K, Introduction to Biometrics System,
          0.4986         1.1164           2.1480                        http://biometrics.cse.msu.edu/.
          0.5386         1.1164           2.2881                   [11] Kumar A., David C.M.Wong, Helen C.Shen, Anil
                                                                        K.Jain, 2004, “Personal Verification using Palmprint
     Figure 6 (a) shows the probability distributions of a              and Hand Geometry Biometric”,
genuine and imposter parts with tolerance value = 3, and                http:/biometrics.cse.msu.edu/Kumar_AVBPA2003.pdf
feature vector length = 256 (16 x 16 blocks). The genuine          [12] LI Wen-xin, David Z,, Shuo-qun XU., 2002,
and imposter parts are estimated from correct and incorrect             Palmprint Recognition Based on Fourier Transform,
matching scores, respectively. The result with various                  Journal of Software, Vol.13, No.5
threshold and false acceptance rates (FAR)/false rejection         [13] Pang Y., Andrew T.B.J., David N.C.L., Hiew Fu San.,
rates (FRR) are shown in figure 6 (b). The equal error rate             2003, Palmprint Verification with Moments, Journal of
(EER) of the verification system is 1.2758. Table 2 show                WSCG, Vol.12, No.1-3, ISSN 1213-6972, Science
the performance (FAR/FRR) system with some threshold                    Press.
values.                                                            [14] Sarraille, J., 2002, Developing Algorithms For
     The main advantage by using PIFS code in this paper                Measuring Fractal Dimension, http://ishi.csustan.edu
is both palmprint feature and palmprint image can be               [15] Shu W., Zhang D., 1998, Automated personal
obtained directly from compressed domain (fractal code).                identification by palmprint, Opt. eng., Vol. 37, No.8,
                                                                        pp. 2359-2363.
                                                                   [16] Tao Y., Thomas R.I., Yuan Y.T., Extraction of
      VII. CONCLUSIONS AND FUTURE WORK                                  Rotation Invariant Signature Based On Fractal
                                                                        Geometry, http://cs.tamu.edu
         In this paper, we introduced a fractal
characteristics based feature extraction and representation

                                                              51                            http://sites.google.com/site/ijcsis/
                                                                                            ISSN 1947-5500
                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                 Vol. 9, No. 2, February 2011

[17] Wohlberg B., Gerhanrd de Jager, 1999, A Review of
    the Fractal Image Coding Literature, IEE
    Transactions on Image Processing, Vol. 8, No.12.
[18] WU Xiang-Quan, Kuan-Quan Wang, David Zhang,
    2004, An Approach to Line Feature Representation
    and Matching for Palmprint Recognition, Journal of
    Software, Vol.15., No.6.
[19] Yokoyama T., Sugawara K., Watanabe T., Similarity-
    based image retrieval system using partitioned
    iterated function system codes, The 8th International
    Symposium on Artificial Life and Robotics, January
    24-26            2006,            Oita,          Japan,
[20] Yokoyama T., Watanabe T., Koga H.,Similarity-
    Based Retrieval Method for Fractal Coded Images in
    the Compressed Data Domain,
[21] Zhang D., Wai-Kin Kong, Jane You, Michael Wong,
    2003, Online Palmprint Identification, IEEE
    Transaction on Pattern Analysis and Machine
    Intelligence, Vol.25, No.9.
[22] Zhang D., and W.Shu, Two novel characteritics in
    palmprint verification: datum point invariance and
    line feature matching, pattern recognition vol 32,

                  AUTHOR PROFILE

                      Dr. I Ketut Gede Darma Putra is a
lecturer in Department of Electrical Engineering and
Information Technology, Udayana University Bali,
Indonesia. He obtained his master and doctorate degree on
informatics engineering from Electrical Engineering,
Gadjah Mada University, Indonesia. His research interest
includes biometrics, image processing, expert system and
Soft computing.

                                                              52                            http://sites.google.com/site/ijcsis/
                                                                                            ISSN 1947-5500

To top