Competitive Coding Scheme for Palmprint Verification

1,2
Adams Wai-Kin Kong and 1David Zhang
1
Biometrics Research Center, Department of Computing,
The Hong Kong Polytechnic University, Kowloon, Hong Kong
[email protected]
2
Pattern Analysis and Machine Intelligence Lab, University of Waterloo,
Waterloo, Ontario, N2L 3G1, Canada
[email protected]

Abstract technologies, palmprint research is just in its infancy. Few

papers have so far been published in this area [2-6].
There is increasing interest in the development of
reliable, rapid and non-intrusive security control systems. A palmprint image contains various features, including
Among the many approaches, biometrics such as principal lines, wrinkles, ridges, minutiae points, singular
palmprints provide highly effective automatic mechanisms points and texture. Lines and texture are the most clearly
for use in personal identification. This paper presents a observable features in low-resolution palmprint images
new method for extracting features from palmprints using (such as 100 dpi). Lines are more appealing than texture
the Competitive Coding Scheme and angular matching. for the human vision. When human beings compare two
The Competitive Coding Scheme uses multiple 2-D Gabor palmprint images, they instinctively compare line features.
filters to extract orientation information from palm lines. This action motivates us to develop a coding scheme for
This information is then stored in a feature vector called the palm lines.
the Competitive Code. The angular matching with an A line contains various information, including 1) type,
effective implementation is then defined for comparing the 2) width, 3) position, 4) magnitude and 5) orientation.
proposed codes, which can make over 9,000 comparisons There are two types of lines: positive and negative [8]. All
within 1s. In our testing database of 7,752 palmprint the lines in our palmprint images are of the negative lines
samples from 386 palms, we can achieve a high genuine category. Palm lines do have a certain width. Generally,
acceptance rate of 98.4% and a low false acceptance rate principal lines are wider than wrinkles but this information
of 3ಎ10-6%. The execution time for the whole process of was not regarded as a useful feature in the previous
verification, including preprocessing, feature extraction palmprint research. On top of types and width, line
and final matching, is 1s. position is often considered as an important feature,
especially for line-based approaches [2]. Magnitude of the
lines has also been investigated in a previous palmprint
1. Introduction
research [5]. It is worthwhile to note that no previous
palmprint research has investigated the orientation
Biometric personal identification is emerging as a
information of the palm lines for palmprint
powerful means for automatically recognizing identities. It
identification/verification. This paper is the first attempt
concerns with identifying people by their physiological
to fully utilize this information for palmprint verification.
characteristics such as iris pattern, retina, palmprint,
In the paper, we only consider feature extraction and
fingerprint, hand geometry and face or some behavioral
matching. The detailed information about palmprint
aspects such as voice, signature and gesture [1].
capture and preprocessing can be referred [6]. The rest of
Automatic fingerprint personal identification has drawn
this paper is organized in the following sections. Section 2
considerable attention over the past 25 years. Recently,
highlights the design, properties and implementation of
voice, face and iris verification/identification have been
the proposed coding scheme. Section 3 describes the
studied extensively. Consequently, many biometric
angular matching for comparing the codes and its
systems have been developed successfully for various
effective implementation. Section 4 reports experimental
applications, including airport security control, ATM
results. Section 5 summarizes the main results of this
(automatic teller machine) and access control [1].
paper and offers concluding remarks.
Nevertheless, compared with other biometric

Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04)

1051-4651/04 $ 20.00 IEEE
 ­° −κ κ2
2
2. Coding scheme design A 8πκσL g − ½° (3)
R(x, y,φ,ω,κ ,σ L ) = − ®e 2 −e 2
¾
ω2σ L2 + κ 2 (1 + 3sin2 φ) °̄ °¿
2.1. Neurophysiology-based Gabor function

Some tunable filters are appropriate for capturing the § ·

¨ κ 2 cos 2 φ ¸
orientation information from palm lines. Gabor filters are where Ø=șˀșL and g= ¨ 1− 2 2 2 .
©ω σ +κ (1+ 3 sin 2 φ ) ¸
L ¹
a good choice. Based on the neurophysiological evidence
from the visual cortex of mammalian brains and wavelet According to Eq. 3, we can obtain the following
theory, Lee reformed the Gabor functions as the following properties.
form [9]: Property 1: R(x,y,Ø,ω,ț,σL) reaches minimum when Þ=0
ω2 Property 2: R(x,y,Ø,ω,ț,σL) is an increasing function with
§ iωx' −κ
2
ω −
2 2
( 4 x' + y ' ) ·
ψ ( x, y , ω , θ ) = e 8κ ¨e − e 2 ¸ respect to ØG when 0<Þcπ/2.
2
(1)
2π κ ¨
©
¸
¹ Property 3: R(x,y,Ø,ω,ț,σL) is a symmetry function with
respect to Ø.
where x’=(x-x0)cosș+(y-y0)sinș, y’=-(x-x0)sinș+(y-y0)cosș; (x0,
Property 4: R(x,y,Ø,ω,ț,σL) is proportional to A, the
y0) is the center of the function; ω is the radial frequency
magnitude of the line.
in radians per unit length and ș is the orientation of the
Property 5: R(x,y,Ø,ω,ț,σL) is independent of C, the
Gabor functions in radians. The ț is defined by
brightness of the line.
§ 2δ + 1 · , where į is the half-amplitude
κ = 2 ln 2 ¨ ¸
¨ 2δ − 1 ¸ Property 6: R(x,y,Ø,ω,ț,σL)=0 when the orientation the
© ¹ filter is perpendicular to the orientation of the line.
bandwidth of the frequency response, which, according to
neurophysiological findings, is between 1 and 1.5 octaves The brightness of the line, C, is removed by the zero
[9]. When ı and į are fixed, Ȧ can be derived from DC Gabor filters. However, according to Property 4, the
ω = κ / σ . These neurophysiology-based Gabor functions response is sensitive to the contrast of the capture devices.
are the same as the general Gabor functions but the Our goal is to design feature codes, which are completely
choices of parameters are limited by neurophysiological independent of the contrast and the brightness of the
findings and the DC (direct current) of the functions are capture devices. The feature codes holding these two
removed. These Gabor functions will be used to extract properties are more robust to different capturing
orientation information from palm lines. environments and devices. Thus, we do not directly use
the response. Based on the six properties, we can design a
competitive rule to extract the orientation information on a
2.2. Competitive rule: winner-take-all
palm line. Our competitive rule is a Winner-take-all rule
defined as: arg minj(I(x,y)*ψR(x,y,ω,șj)) where I is a
To design an explainable competitive rule for
preprocessed image; ψR represents the real part of ψ ; șj
extracting orientation information from palm lines, we
is the orientation of the filters and j={0,…,J}. According
have to construct an idea palm line model, whose profile
to the neurophysiological findings, the simple cells are
has an upside-down Gaussian shape. It is given by:
sensitive to specific orientations with approximate
ª § (( x − x p ) cos θ L + ( y − y p ) sin θ L ) 2 ·º
L ( x , y ) = A «1 − exp ¨ − ¸» + C (2) bandwidths of ʌ/6 [9]. Thus, we choose six orientations,
«¬ ¨ 2σ L2 ¸»
© ¹¼
șj=jπ/6, where j={0, 1,…,5} for the competition.
where σL, the standard deviation of the profile, can be If we only extract the orientation information on the
considered as the width of the line; (xp, yp) is the center of palm lines, we have to face two problems:
the line; A, a positive real number, controls the magnitude 1) How do we classify a point that belongs to a palm line?
of the line, which depends on the contrast of the capture 2) Even though we can have a good technique to classify
device; C is the brightness of the line, which replies on the points on the palm lines, to design a real-time
brightness of the capture device and the lighting of the classifier for the feature points for identifying a palm in a
capture environment and șL is the orientation of the line. large database is difficult - the number of the extracted
Without loss of generality, we set xp=0 and yp =0 for the feature points may be different, even for two palmprint
following analysis. images belonging to the same palm.
To extract the orientation information from the palm To avoid these two problems and to achieve our goal
lines, we apply the real part of the neurophysiology-based Ё real-time palmprint verification – we assume that each
Gabor filters to the idea palm line model. The filter
point on the palmprint belongs to a palm line. Thus, we
response on the middle of the line, xcosșL+ysinșL=0, is:
apply our competitive rule to code each sample point to
obtain feature vectors with the same dimension. The
proposed feature is named as Competitive Code. An

Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04)

1051-4651/04 $ 20.00 IEEE
 example of Competitive Code is illustrated in Fig. 1. The needs 0.11ms for one matching. This bit representation is
codes of 0-5 are shown in Figs. 1(c)-(h), respectively. not only effective for matching but also effective for
Carefully comparing Fig. 1(a) and Fig. 1(b) (a storage. In total, three bits are enough to keep the mask
combination of Figs. 1(c)-(h)), we can observe that the and one element of the Competitive Code. If a non-
Competitive Code is highly related to the line features, palmprint pixel exits at position (x,y), the corresponding
especially for the strong lines, such as the principal lines. three bits are set to 1, 0 and 1. As a result, the total size of
the proposed feature, including the mask and the
3. Angular matching Competitive Code is 384 bytes.

To implement a real-time palmprint identification

system requires a simple and powerful palmprint matching
algorithm. Since the proposed Competitive Code stores
the orientation information, we design an angular distance
for comparing two codes. Let P and Q be two Competitive
Codes and PM and QM be the corresponding masks of P (a) (b)
and Q, respectively. The masks are used to indicate the
non-palmprint pixels described in [6]. The angular
distance is defined as:
N N

¦¦ ( P M ( x, y ) ∩ Q M ( x, y )) × G ( P ( x, y ), Q ( x, y ))
y =0 x = 0
D ( P, Q ) = N N
(4)
3¦¦ PM ( x, y ) ∩ QM ( x, y ) (c) (d) (e)
y =0 x=0

where
­min(P(x, y) −Q(x, y), Q(x, y) −(P(x, y) −6)) if P(x, y ≥ Q(x, y)
G(P(x, y),Q(x, y)) = ®
¯min(Q(x, y) − P(x, y),P(x, y) −(Q(x, y) −6)) if Q(x, y) > P(x, y)
(5)
and ∩ represents an AND operator and the size of the
feature matrixes is N×N. Obviously, D is between 0 and 1. (f) (g) (h)
For perfect matching, the angular distance is zero.
Because of imperfect preprocessing, we need to translate Figure 1. An example of Competitive Code. (a)
one of the features vertically and horizontally and then Preprocessed image. (b) Competitive code. (c)-(h) are the
perform the matching again. Both the ranges of the winning Code 0, 1, 2, 3, 4 and 5, respectively.
vertical and the horizontal translation are –2 to 2. The
minimum of the D’s obtained by translated matching is Table 1. Bit representation of the Competitive Code
regarded as the final angular distance.
However, we find that directly implementing Eqs. 4-5 Original values Bit 1 Bit 2 Bit 3
is ineffective. The elements of Competitive Code are 0, 1, 0 0 0 0
2, 3, 4 and 5. We can use three bits to represent an 1 0 0 1
element and one bit for the mask. In total, a Competitive 2 0 1 1
Code is constituted by four bit-planes. The bit values 3 1 1 1
among different elements of Competitive Code are shown 4 1 1 0
in Table 1. According to this bit representation of the 5 1 0 0
Competitive Code, a more effective implementation of
angular distance can be defined as:
4. Experimental results
N N 3

¦ ¦ ¦ (P M ( x, y ) ∩ Q M ( x , y )) ∩ ( Pi ( x, y ) ⊗ Q ( x, y ))
b
i
b

D ( P, Q ) =
y =0 x =0 i=0
We collected palmprint images from 193 individuals
N N
3¦ ¦ PM ( x, y ) ∩ Q M ( x, y )
(6)
using our palmprint scanner. The subjects are mainly
y =0 x =0
students and staff volunteers from The Hong Kong
where Pi (Q ) is the ith bit plane of P(Q) and ⊗ is bitwise
b
i
b
Polytechnic University. In the dataset, 131 people are
exclusive OR. male, and the age distribution of the subjects is: about
Using ASUS notebook embedded Intel Pentium III 86% are younger than 30, about 3% are older than 50, and
Mobile processor (933MHz), directly implementing Eqs. about 11% are aged between 30 and 50. We collected the
4-5 requires 2.27ms for one matching but Eq. 6 only palmprint images in two occasions. Each time, the

Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04)

1051-4651/04 $ 20.00 IEEE
 subjects were asked to provide around 10 images from the 98.4% and a low false acceptance rate of 3ಎ10-6%. In the
left palm and 10 images from the right palm. Altogether, experiments, we have compared the Competitive Code
each person provided around 40 images, resulting in a with the PalmCode, which is our previously best approach
total of 7,752 images from 386 different palms in our [6]. It is no doubt that Competitive Code is more accurate
database. The average time interval between the first and than PalmCode.
the second collection is 69 days. The maximum and the
minimum time intervals are 162 and 4 days, respectively.

4.1 Verification

To obtain the verification accuracy of our method, each

palmprint image is matched with all the other palmprint
images in the database. A match is counted as correct if
the two palmprint images are from the same palm;
otherwise, the match is counted as incorrect. The total
number of matches is 30,042,876. None of the angular
distances is zero. The number of comparisons that match Figure 2 The receiver operator characteristic (ROC)
correctly is 74,068. The rest are incorrect. Fig. 2 depicts curves of the Competitive Code and the PalmCode.
the corresponding Receiver Operating Characteristic
(ROC) curve as a plot of the genuine acceptance rate
against the false acceptance rate for all possible operating
References
points. In Fig. 2 we can see that the proposed method can
[1] A. Jain, R. Bolle and S. Pankanti (eds.), Biometrics:
operate at genuine acceptance rate of 98.4% while the Personal Identification in Networked Society, Boston,
corresponding false acceptance rate is 3ಎ10-6%. In Fig. Mass: Kluwer Academic Publishers, 1999.
2, the other ROC curve is obtained by our previously best [2] D. Zhang and W. Shu, “Two novel characteristics in
method, PalmCode [6]. Comparing these two ROC palmprint verification: datum point invariance and line
curves, we ensure that the proposed Competitive Code is feature matching,” Pattern Recognition, vol. 32, no. 4, pp.
much better than PalmCode. The genuine acceptance rate 691-702, 1999.
of Competitive Code is about 13.5% higher than that of [3] N. Duta, A.K. Jain, and K.V. Mardia, “Matching of
palmprint,” Pattern Recognition Letters, vol. 23, no. 4, pp.
PalmCode while their false acceptance rates are 3ಎ10- 477-485, 2001.
6
%. The verification results of Competitive Code are [4] W. Shu, G. Rong, Z. Bain and D. Zhang, “Automatic
comparable with the previous palmprint approaches and palmprint verification,” International Journal of Image and
other hand-based biometric technologies, including hand Graphics, vol. 1, no. 1, pp. 135-152, 2001.
geometry and fingerprint verification [2-7, 10]. [5] C.C. Han, H.L. Cheng, K.C. Fan and C.L. Lin, “Personal
authentication using palmprint features,” Pattern
Recognition, vol. 36, no 2, pp. 371-381, 2003.
5. Conclusions [6] D. Zhang, W.K. Kong, J. You and M. Wong, “On-line
palmprint identification,” IEEE Trans. on PAMI, vol. 25,
We have presented a novel feature extraction method, no. 9, pp. 1041-1050, 2003.
the Competitive Coding Scheme for palmprint [7] D. Maio, D. Maltoni R. Cappelli, J.L. Wayman and A.K.
identification. This scheme extracts the orientation Jain, “FVC 2000:Fingerprint verification competition,”
information from the palm lines and stores it in the IEEE Trans. on PAMI, vol. 24, no 3, pp. 402-412, 2002.
[8] J.H. Van Deemter and J.M.H. Du Buf, “Simultaneous
Competitive Code. An angular match with an effective
detection of lines and edges using compound Gabor filters,”
implementation is developed for comparing Competitive International Journal of Pattern Recognition and Artificial
Codes. Using an ASUS notebook embedded Intel Pentium Intelligence, vol. 14, no. 6, pp. 757-777, 2000.
III Mobile processor (933MHz) and the bit representation [9] T.S. Lee, “Image representation using 2D Gabor wavelet,”
of the Competitive Code, angular matching can make over IEEE Trans. on PAMI, vol. 18, no. 10, pp. 959-971, 1996.
9,000 comparisons in about 1s. Total execution time for [10] R. Sanchez-Reillo, C. Sanchez-Avilla and A. Gonzalez-
verification is about 1s, which is fast enough for real-time Marcos, “Biometric identification through hand geometry
applications. The proposed coding scheme has been measurements,” IEEE Trans. on PAMI, vol. 22, no 10,
evaluated using a database with 7,752 palmprint images pp.1168-1171, 2000.
from 386 different palms. For verification, the proposed
method can operate at a high genuine acceptance rate of

Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04)

1051-4651/04 $ 20.00 IEEE