Fourier Spectral of PalmCode as Descriptor for Palmprint Recognition

Meiru Mu

1,2

, Qiuqi Ruan

, Luuk Spreeuwers

and Raymond Veldhuis

Institute of Information Science, Beijing Jiaotong University, Beijing, China

Signals and Systems group, University of Twente, Enschede, The Netherlands

Keywords:

Gabor Filtering, Palm Code, Fourier Spectral, Horizontal and Vertical 2DPCA, Palmprint Recognition.

Abstract:

Study on automatic person recognition by palmprint is currently a hot topic. In this paper, we propose a

novel palmprint recognition method by transforming the typical palmprint phase code feature into its Fourier

frequency domain. The resulting real-valued Fourier spectral features are further processed by horizontal

and vertical 2DPCA method, which proves highly efﬁcient in terms of computational complexity, storage

requirement and recognition accuracy. This paper also gives a contrast study on palm code and competitive

code under the proposed feature extraction framework. Besides, experimental results on the Hongkong PolyU

Palmprint database demonstrate that the proposed method outperforms many currently reported local Gabor

pattern approaches for palmprint recognition.

1 INTRODUCTION

Nowadays, recognition of individuals by means of

biometric characteristics is becoming increasingly fa-

miliar and accepted. The selection of biometrics is

commonly application-dependent (D. Zhang, 2004;

S. Pankanti et al., 2000). Palmprint, extracted from

person’s hand, has been recognized as a means of

measurement that can uniquely represent a person

(W. Shu et al., 1998b). It has been a long history

since the palmprints found on the crime scene were

used for forensic investigation. Recently, palmprint

is being investigated intensively for personal recog-

nition in different real-time application system, such

as access control, network security, and social secu-

rity. Compared with the ﬁngerprints, palmprints have

more rich features which are less likely to be de-

stroyed and forged. Besides, palmprints can be cap-

tured with a much lower resolution imaging sensor

(less than 100 dpi), which leads to be more efﬁcient

(W. Shu et al., 1998a). The online palmprint capture

devices are mainly based on CCD camera or digi-

tal scanner. More recently, the real time multispec-

tral palmprint capture device has also been developed

(R.K. Rowe et al., 2007; Z. Guo et al., 2010). In

addition to the efﬁcient palmprint acquisition, robust

palmprint representation is another key issue for the

success of palmprint recognition application.

The algorithms proposed for online palmprint

recognition (including veriﬁcation and identiﬁcation),

are generally divided into three main classes: sub-

space learning methods, texture energy feature extrac-

tion, and coding based methods. Among them, coding

based methods are deemed to be the most promising

due to their high recognition accuracy and small fea-

ture size, which typically involve steps of ﬁlter bank

selection, coding scheme design and template match-

ing approach. The popular ﬁlters include Gabor,

Gaussian, and other self-designed ones. The phase,

orientation, and magnitude information are generally

regarded as encoded objects. The coding rules are

usually ﬂexible and simple, and the obtained code fea-

tures should be robust and provide high discriminative

ability. PalmCode (D. Zhang et al., 2003) encodes the

phase of Gabor ﬁltered responses into binary features.

FusionCode (A. Kong et al., 2006) used a fusion rule

at feature layer to further improve PalmCode. DoG

code method (X.Q. Wu et al., 2006) ﬁrst convolves

the image using two-dimensional Gaussian ﬁlter and

then encodes the zero-crossing information of hori-

zontal and vertical gradient values, respectively. Or-

dinalCode (Z.N. Sun et al., 2005) encodes the outputs

of the elliptical Gaussian ﬁlters. The schemes encod-

ing the orientation of palm lines are also popular due

to its stability which includes competitive code (Com-

pCode) (A. Kong et al., 2004), binary orientation co-

occurrence vector (BOCV) code (Z. Guo et al., 2009),

palmprint orientation code (POC) (X.Q. Wu et al.,

2005), robust line orientation code (RLOC) (W. Jia

et al., 2008), and so on (F. Yue et al., 2009; W. Zuo

Mu M., Ruan Q., Spreeuwers L. and Veldhuis R. (2013).

Fourier Spectral of PalmCode as Descriptor for Palmprint Recognition.

In Proceedings of the 2nd International Conference on Pattern Recognition Applications and Methods, pages 22-32

DOI: 10.5220/0004228000220032

 SciTePress

et al., 2010). However, coding based methods re-

quire the pixel-to-pixel paired matching between en-

rollment and query templates. In order to offset the

inter-class translation variation, one template has to

be shifted in some range and the matching score com-

putation is proceeded multiple times to ﬁnally deter-

mine the identity of one query palmprint, which is

time costly. Refer to Ref. (D. Zhang, 2004), the

matching speed of PalmCode is 1.7 ms when the sys-

tem works under the veriﬁcation mode and the algo-

rithm is implemented in Visual C++ 6.0 on a PC using

Intel Pentium III processor (500MHZ). For identiﬁ-

cation, if the database contains 100 persons and three

palmprint images are registered for each subject, the

total identiﬁcation time is about 1.1s. As can be seen,

the matching speed of coding based methods is not

desirable for the identiﬁcation system with large reg-

istered database. This issue has also been addressed

in some other reported literatures. In order to deal

with it, Jia (W. Jia et al., 2008) applied Tensor LPP on

their extracted directional representation. However, it

can not provide translation invariant performance by

reducing the dimension of integer codes directly. Yue

(F. Yue et al., 2011) proposed an approach to speed

up the coding based palmprint recognition method by

building a template tree to perform fast nearest neigh-

bor searching. For both of these methods, the ﬁnal

classiﬁcation is still based on pixel-to-pixel code tem-

plate matching by Hamming distance.

Accordingly, in this paper, we aim to further study

the code features for palmprint representation. The

following three issues are mainly considered:

Firstly, about the ﬁlter selection, Gabor, Gaussian,

and the second derivative of Gaussian ﬁlter have been

evaluated based on several coding based methods in

Ref. (F. Yue et al., 2008) and their experimental re-

sults show that the Gabor ﬁlter is superior. Besides, it

is well known that Gabor representation can provide

optimal localization of image details in a joint spatial

and frequency domain (J. Beck, 1987). Concerning

these conclusions, the Gabor ﬁlters are chosen for our

proposed feature extraction method in this paper.

Secondly, about the encoded object, we employ

the phase information. Gabor phase, as a discrimi-

nating information source, has been successfully used

in pattern recognition ﬁeld. Besides PalmCode and

IrisCode, kinds of local Gabor phase pattern opera-

tors (W. Zhang et al., 2009; B. Zhang et al., 2007;

S. Xie et al., 2010) are proposed as well based on the

combination of the spatial histogram and the Gabor

phase pattern encoding scheme. These methods in-

volve a group of Gabor ﬁlters (in general 40 Gabor

ﬁlters) to extract the multi-scale and multi-orientation

information, which is an overcomplete representation

with a high redundant ratio. The high dimensional-

ity of the local Gabor phase pattern histograms leads

to high computational cost and large storage require-

ment. To deal with this problem, this paper proposes a

new method, which uses only two Gabor ﬁlters while

provides higher recognition performance.

Thirdly, instead of representing palmprint image

by the code features directly, we transform the code

feature matrices into their Fourier frequency ﬁelds

since a shift in the time domain causes no change

in the Fourier magnitude spectrum. The resulting

spectral features not only keep the property of pow-

erful discriminative ability of PalmCode features, but

also make it possible to extend the application ﬁeld

of coding based features. Because the dissimilarity

measurement of the original coding based features is

generally based on pixel-to-pixel matching by using

Hamming or angular distances, which limits their ap-

plication to unsupervised template matching mode,

while the real-valued spectral features might be pro-

cessed further by the currently reported supervised

learning algorithms, which probably achieves higher

recognition performance.

PalmCode DFT

Horizontal

2DPCA

Vertical

2DPCA

Palm

Figure 1: Flow chart of the proposed method. (DFT - dis-

crete Fourier transform; 2DPCA - two-dimensional princi-

ple components analysis).

Taking all the above factors into consideration,

a novel Fourier spectral representation of PalmCode

(DFT PalmCode) is proposed for palmprint recog-

nition in this paper, whose ﬂow chart is shown in

Fig. 1. We ﬁrstly transform the typical palmprint

phase code feature into its Fourier frequency domain

by discrete Fourier transform (DFT). The resulting

real-valued Fourier spectral features are then pro-

cessed by the horizontal and vertical two-dimensional

principal component analysis ((2D)

PCA) method.

The translation invariant property of Fourier trans-

form and (2D)

PCA dimensionality reduction method

help alleviating the within-class image translations

to some extent. Compared with the original Palm-

Code method, the proposed feature extraction ap-

proach operates two pairs of Gabor ﬁltering phase

responses, which improves the recognition perfor-

mance evidently. This paper also gives a contrast

study on CompCode under the proposed feature ex-

traction framework. For CompCode, palm line orien-

tation information is encoded into bits. Furthermore,

our experimental results demonstrate that the pro-

posed method outperforms many currently reported

local Gabor pattern operators for palmprint recogni-

tion by higher accuracy, lower computational cost and

FourierSpectralofPalmCodeasDescriptorforPalmprintRecognition

less storage requirement.

This paper is organized as follows. In section 2,

we give a short review of PalmCode. Section 3 il-

lustrates the proposed Fourier spectral representation

in details. In section 4, the (2D)

PCA method is de-

scribed for the dimensional reduction of the proposed

spectral features, which helps alleviating the negative

effects due to image translation. The experimental re-

sults will be demonstrated in section 5. The ﬁnal part

is about the conclusions.

2 REVIEW OF PalmCode

The Gabor phase quadrant demodulation coding

method is proposed by Daugman for iris recogni-

tion in 1993 (J. Daugman, 1993), which is then

reported successful for palmprint representation in

Ref. (W.K. Kong et al., 2003). For ease of presen-

tation, this feature extraction method is commonly

named by IrisCode, PalmCode or Daugman’s method.

For PalmCode, a circular 2-D Gabor ﬁlter is used

to ﬁrstly convolve the original gray palmprint image,

which has the following general form:

g(x,y,u,θ,σ) =

2πσ

exp

{

−(x

+ y

)

2σ

}

×exp{2πi(uxcosθ + uy sin θ)}.

(1)

Where i =

√

−1, u is the frequency of the sinusoidal

wave along the direction θ from the x-axis, and σ

speciﬁes the Gaussian envelope along x and y axes,

which determines the bandwidth of the Gabor ﬁlter.

In practice, a Gabor function with a special set of

parameters {u, θ, σ}, is transformed into a discrete

Gabor ﬁlter. In order to provide more robustness to

brightness, the Gabor ﬁlter is turned to zero direct

current (DC) denoted by ˜g(x,y). Following the ex-

perimental setting in Ref. (W.K. Kong et al., 2003), σ

is set to 5.6179.

Given an image f (x,y) of size M ×N, its Gabor

ﬁltered images are deﬁned as:

G(x,y) =

∑

f (x

) · ˜g(x −x

,y −y

). (2)

Where G(x,y) is complex number with real part

(G(x,y)) and imaginary part

(G(x,y)).

Assuming the Gabor parameters {θ,u,σ} are

given, PalmCode method encodes each pixel in the

Gabor ﬁltered image into two bits by the following

rules:

(x,y) =

{

if Re

(G(x,y)) < 0;

otherwise

(3)

(x,y) =

{

if Im

(G(x,y)) < 0;

otherwise

(4)

Where P

(x,y) and P

(x,y) are respectively the en-

coded real and imaginary binary feature templates.

3 FOURIER SPECTRAL OF

PalmCode

The discrete Fourier transform (DFT) of a func-

tion (image) p(x, y) of size M ×N can be given by

P(u,v) = |P(u,v)|e

ϕ(u,v)

, where |P(u,v)|= [R

(u,v)+

(u,v)]

1/2

is called the magnitude spectrum of the

Fourier transform, and ϕ(u,v) = tan

−1

[

I(u,v)

R(u,v)

]

called the phase angle or phase spectrum of the trans-

form. R(u,v) and I(u,v) are the real and imaginary

parts of F(u,v), respectively.

The Fourier transform has the following transla-

tion properties:

If p(x, y) ⇔ P(u,v),

then

p(x −x

) ⇔ P(u, v)e

−j2π(ux

/M+vy

/N)

(5)

and

p(x,y)(−1)

x+y

⇔ P(u −M/2,v −N/2) . (6)

Equation (5) tells us that a shift in the time domain

causes no changes in the magnitude spectrum but only

the changes in the phase spectrum of Fourier trans-

form. This property has been widely used to ex-

tract translation invariant features in pattern recogni-

tion. Equation (6) describes that multiplying p(x, y)

by (−1)

x+y

shifts the original of P(u,v) to frequency

coordinates (M/2,N/2), which is the center of the

M ×N area occupied by the 2-D DFT. This result is

based on the variables u and v having values in the

range [0, M −1] and [0,N −1], respectively. In a com-

puter implementation these variables will run from

u = 1 to M and v = 1 to N, in which case the ac-

tual center of the transform will be at u = M/2 + 1

and v = N/2 + 1. This property is usually used when

visualizing the 2-D Fourier spectrum.

The DFT has the following conjugate symmetry

property:

P(u,v) = P

∗

(−u,−v), (7)

from which it follows that the spectrum is symmetric

about the original:

|P(u,v)| = |P(−u,−v)|. (8)

Based on the symmetry property, we can only keep

half of the Fourier coefﬁcients for image representa-

tion.

Figure 2 gives an illustration of the proposed

Fourier spectral representation scheme. Figure 2(a)

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

One Gabor

filter

Real part

Imaginary

part

(a)

PalmCode Fourier spectral

(b) (c) (d)

DFT

(e) (f)

Figure 2: An illustration of the proposed Fourier spectral representation scheme. (a) original image; (b) PalmCode; (c)

PalmCode down-sampled by ratio of 2:1; (d) Palmcode with the four rows and columns along the code plane edges set to 0;

(e) Fourier spectral of (d); (f) half of (e).

shows a typical palmprint image from PolyU palm-

print database. Figure 2(b) visualizes the original

PalmCode feature planes as illustrated in section 2.

For reducing the computational cost, the PalmCode

features are down-sampled by the ratio of 2:1 in this

study, as can be seen from Fig. 2(c). On the edge

of the code matrices, there are usually some non-

palmprint features resulted from defective image pre-

processing. For reducing their affects, some masks

are generally needed in the classical coding based

methods (D. Zhang et al., 2003; A. Kong et al., 2004).

While in the proposed method, the values of four rows

and columns along the code plane edges are changed

into zeros for removing the non-palmprint pixels, as

Fig. 2(d) shows. Figure 2(e) shows the Fourier spec-

tral features of PalmCode with the center coordinate

(33,33). It is quite obvious that the Fourier spectral

matrix is symmetric with respect to the center point.

Thus, only half of the spectral values are required for

further processing. Figure 2(f) shows half of the spec-

tral features, which is the ﬁnal Fourier spectral repre-

sentation of our proposed.

4 SPECTRAL FEATURE

REDUCTION BY (2D)

PCA

The Fourier spectral feature of PalmCode consists

of several real-valued matrices size of 31×64, as

shown in Fig. 2(f). The large dimensionality leads

to high computational cost, limited matching speed

and costly template storage requirement. In order to

cope with these problems, we resort to (2D)

PCA di-

mensional reduction technique (X. Pan et al., 2008),

which is called Modiﬁed 2DPCA (D. Zhang et al.,

2009) or horizontal and vertical 2DPCA method

(J. Yang et al., 2007) in some other papers. The

(2D)

PCA method not only helps saving the com-

putational cost and providing smaller feature space,

but also has a useful invariance property: the trans-

form matrix of horizontal 2DPCA is invariant to any

change of image row sequence, and the transform ma-

trix of vertical 2DPCA is invariant to any change of

an image column sequence. This invariance property

of (2D)

PCA helps alleviating the negative effect re-

sulting from image translation (J. Yang et al., 2007).

Taking these factors into consideration, (2D)

PCA

method is sequently carried out on the Fourier spec-

tral feature space.

Given N spectral feature X

,...,X

, each spec-

tral X

(i = 1,2,...,N) is a m ×n matrix, the goal

of horizontal 2DPCA is to ﬁnd the optimal orthog-

onal projection axes U so that the projected matrices

Y = [Y

,...,Y

] achieve a maximum total scatter,

which can be expressed in the form:

= X

U, i = 1, 2, . . . , N (9a)

U = arg(max(tr(S

))), (9b)

where

tr(S

) = U

U (10a)

∑

j=1

−

Y )(Y

−

Y )

(10b)

∑

j=1

−

X) , (10c)

It follows that S

is a non-negative deﬁnite matrix

size of n × n. In general, the optimal projection

axes U = [u

,...,u

] are chosen as the orthogonal

eigenvectors of S

corresponding to the largest eigen-

values λ

,λ

,...,λ

. Thus the projected vectors are

d-dimensional. Finally, after processed by the hori-

zontal 2DPCA, each spectral matrix is represented as

a matrix Y

of dimension m ×d (d ≪ n).

After the procedure of horizontal 2DPCA trans-

form, the correlation between row vectors of spectral

feature matrices is removed, and the resulted projec-

tion matrices are less sensitive to the translation and

FourierSpectralofPalmCodeasDescriptorforPalmprintRecognition

mirror variations of the spectral row sequences. Fur-

ther, the vertical 2DPCA transform is needed to re-

move the correlation between column vectors of the

projected matrices Y

For simplify, the horizontal 2DPCA transform is

processed on the transpose of the projected matrices,

denoted by Y

(i = 1,2,...,N). Then the correspond-

ing total scatter matrix can be expressed in the form:

∑

j=1

−

)

−

)

∑

j=1

−

Y )(Y

−

Y )

(11)

which is a non-negative deﬁnite matrix size of m ×m.

Let V = [v

,...,v

] be the orthogonal eigenvec-

tors of C

corresponding to the largest eigenvalues

,γ

,...,γ

(q ≪m). By letting R represent the pro-

jected matrices of Y after the vertical 2DPCA trans-

form, we have R

is the projected matrices of Y

after

the horizontal 2DPCA transform, that is

= Y

V . (12)

It follows that

R = (Y

V )

= V

Y = V

XU, R ∈ R

q×d

. (13)

As can be seen, the dimensionality of the Fourier

spectral feature X is reduced from m × n to q ×d

(d ≪ n, q ≪ m) after the (2D)

PCA transform.

As Fig. 2 shows, each image generates two spec-

tral feature matrices when one Gabor ﬁlter is con-

volved. Based on the proposed feature extraction

framework, we might obtain more robust feature rep-

resentation for palmprint recognition, if more Gabor

ﬁlters with different orientations are used. As a mat-

ter of fact, we ﬁnd that the proposed method performs

much better when two Gabor ﬁlters instead of one are

used for spectral feature extraction according to the

experimental results. Figure 3 illustrates the training

steps of the proposed algorithm when two Gabor ﬁl-

ters are used, in accordance with which we summarize

the proposed palmprint feature extraction algorithm

as follows:

Step 1. Input the training palmprint images, and con-

volve each image with two Gabor ﬁlters of

different orientation parameters to obtain four

PalmCode matrices.

Step 2. Perform the 2-D Fourier transform on each

binary PalmCode matrix to get the corre-

sponding spectral feature matrix as described

in Fig. 2.

Step 3. Perform (2D)

PCA transform on these spec-

tral feature matrices got in Step 2 to obtain the

Gabor

filter1

(2D)

PCA

Real part1

Gabor

filter2

Training

samples

Imaginary

part1

Real part2

Imaginary

part2

and

Fourier spectral

of PalmCode

Projection

of matrices

(2D)

PCA

(2D)

PCA

(2D)

PCA

Figure 3: The training steps of the proposed algorithm when

two Gabor ﬁlters are used. The Fourier spectral of Palm-

Code is achieved as illustrated in Fig. 2.

corresponding four groups of feature projec-

tion matrices (V 1,U1), (V 2,U2), (V 3,U3),

(V 4,U4) as shown in Fig. 3.

Step 4. Transform the four Fourier spectral matrices

of each training image by the four groups

of feature projection matrices respectively to

generate four feature matrices as described in

Eq. (13), denoted by {R1,R2,R3,R4}. Save

them as the training feature database.

Step 5. When a query image is captured, the same

processing as Steps 1 and 2 is applied, and

then transform their Fourier spectral matri-

ces as Step 4 to get the ﬁnal feature template

{R1

′

,R2

′

,R3

′

,R4

′

Step 6. For template matching, the training and

test feature matrices are respectively con-

verted into vectors {R1,R2, R3, R4} and

{R1

′

,R2

′

,R3

′

,R4

′

}; Then the dissimilarity is

measured by sum of Euclidean distance be-

tween each pair of vectors Ri and Ri

′

, which

can be denoted by

∑

i=1

D(Ri,Ri

′

1(a) 2(a) 3(a)

1(b) 2(b) 3(b)

4(a) 5(a)

4(b) 5(b)

Figure 4: The palmprint images in the Hong Kong PolyU

database. (a) Samples in session one; (b) Samples in session

two.

5 EXPERIMENTAL RESULTS

In this section, we evaluate the recognition perfor-

mance of our proposed method using HongKong

Polytechnic University (PolyU) Palmprint Database,

which contains 7752 grayscale images in BMP image

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

2 4 6 8 10

96.5

97.5

98.5

99.5

100

Ten cases with different Gabor orientation settings

Identification accuracy (%)

One Gabor filter

Two Gabor filters

Three Gabor filters

Figure 5: The identiﬁcation accuracies (%) versus the ten

Gabor orientation setting cases.

format. They were captured by CCD camera from

386 different palms and collected in two occasions

with different illumination conditions. The interval

between the two sessions is about two months. Each

person provided around ten palmprint images from

the left and right hands respectively. The resolution of

original captured images is 384×284 pixels at 75 dpi.

By performing the similar preprocessing approach de-

scribed in Ref. (W.K. Kong et al., 2003), we cropped

the region of interest (ROI) size of 128×128. Figure 4

shows the cropped images from ﬁve typical palm sam-

ples, in which (a) and (b) denote the samples captured

from the same palm at the ﬁrst and second sessions

respectively.

5.1 Determination of the Parameters

For the Gabor based methods, there is generally a

set of adjustable parameters. For DFT PalmCode

method, Besides Gabor parameters, d and q for

(2D)

PCA transform are controllable as well. In this

group of experiments, we ﬁrstly calculate the values d

and q by taking 90% of the total sum of the eigenval-

ues. Based on the obtained values d and q, a series of

experiments is carried out to investigate the inﬂuence

of Gabor orientation parameter and examine how the

number of Gabor ﬁlters affects the identiﬁcation rate.

Secondly, with the determined optimal Gabor ﬁlters,

the identiﬁcation accuracies are investigated as d and

q vary.

For each palm, ﬁve samples from session one are

randomly collected to construct the training set, and

all the samples (in total 3863) captured in session two

are used for testing. Assuming the system is operated

in the identiﬁcation mode (one-to-many comparison).

The average correctness rate (%) of ten-run identiﬁ-

cation executions is used to evaluate the identiﬁcation

performance.

By following the experimental results of literature

(W.K. Kong et al., 2003), the Gabor parameter u is

set to 0.0916, and σ is ﬁxed to 5.6179. In order to

determine the optimal number of Gabor ﬁlters and

their orientation parameters for our proposed method,

16 18 20 22 24 26 28 30 32 34 36 38

99.45

99.5

99.55

99.6

99.65

99.7

99.75

99.8

The projected vector dimension d of the horizontal 2DPCA

Identification accuracy (%)

q=10

q=12

q=14

q=16

q=18

q=20

q=22

Figure 6: Identiﬁcation rates (%) as values d and q vary.

we test some orientation cases on palmprint identi-

ﬁcation accuracy. For the applied (2D)

PCA trans-

form, values d and q are determined by taking 90%

of the total sum of eigenvalues. When one Gabor

ﬁlter is used, the considered orientation values are

◦

, 10

◦

, 30

◦

, 45

◦

, 60

◦

, 90

◦

, 120

◦

, 135

◦

, 150

◦

, and

170

◦

, respectively. Besides, the identiﬁcation rates

are also investigated when two and three Gabor ﬁl-

ters are used with ten cases of different orientation

settings. The tested orientation groups are as follows:

{45

◦

, 60

◦

}, {45

◦

, 90

◦

}, {45

◦

, 120

◦

}, {45

◦

, 135

◦

{60

◦

, 90

◦

}, {60

◦

, 120

◦

}, {60

◦

, 135

◦

}, {90

◦

, 120

◦

{90

◦

, 135

◦

}, {120

◦

, 135

◦

}, {45

◦

,60

◦

, 90

◦

}, {45

◦

, 120

◦

}, {45

◦

, 60

◦

,135

◦

}, {45

◦

, 90

◦

, 120

◦

}, {45

◦

, 135

◦

}, {45

◦

, 120

◦

, 135

◦

}, {60

◦

, 90

◦

, 120

◦

{60

◦

, 90

◦

, 135

◦

}, {60

◦

, 120

◦

, 135

◦

}, {90

◦

, 120

◦

135

◦

}. Figure 5 gives an illustration of the relation-

ship between the identiﬁcation accuracies and the Ga-

bor orientation parameters. As it shows, the recogni-

tion performances of the proposed method improve

when the Gabor ﬁlter number increases from two to

three in most cases. However, the increase of Gabor

ﬁlter numbers adds to the computational cost in multi-

ple. Based on an overall consideration of the recogni-

tion performance and the computational complexity,

case 6 with two Gabor ﬁlters is determined to be op-

timum. That is to say, our proposed method achieves

the highest identiﬁcation rate 99.74% when two Ga-

bor ﬁlters with the orientation parameters of 60

◦

and

120

◦

respectively are used for the spectral feature ex-

traction on the PolyU palmprint database.

Given the Gabor parameters we got above, the op-

timal values d and q for (2D)

PCA are experimen-

tally investigated by varying q from 10 to 22 and d

from 16 to 38. The identiﬁcation rates correspond-

ing to different d and q are plotted in Fig. 6. As can

be seen, the best identiﬁcation rate of our proposed

method is close to 99.8%. The feature dimension af-

ter (2D)

PCA transform is q ×d, which is expected to

FourierSpectralofPalmCodeasDescriptorforPalmprintRecognition

Table 1: Comparisons of the average identiﬁcation accuracy (%), time cost (s) for identifying per test sample, and the veriﬁ-

cation EER (%).

Method Identiﬁcation accuracy (%) Time cost (s) Veriﬁcation EER (%)

PalmCode 0 75.63 0.067 29.5

PalmCode [-1,1] 95.14 0.644 9.1

PalmCode [-2,2] 98.27 1.66 3.1

DFT PalmCode 99.74 0.123 0.7

CompCode 0 78.14 0.085 24.2

CompCode [-1,1] 96.17 0.920 7.2

CompCode [-2,2] 99.11 2.398 1.9

DFT CompCode 99.29 0.106 0.9

as small as possible while keeping the highest identi-

ﬁcation rate. Therefore, the best identiﬁcation perfor-

mance is obtained when d = 28, q = 12, as can be

seen from Fig. 6.

It should be pointed out that in other experiments

we simply take the sum of the largest d eigenvalues

that determine 90% of the total sum of the eigen-

values when processing the horizontal 2DPCA trans-

form. By the same way, we get the value of q for the

vertical 2DPCA transform.

5.2 Comparisons with Coding Methods

In this section the palmprint identiﬁcation and ver-

iﬁcation performances are investigated when Palm-

Code is replaced by the CompCode, which I denote

by DFT CompCode. For PalmCode, the used Gabor

ﬁlter parameters are given by u = 0.0916, θ = π/4,

and σ = 5.6179; For CompCode, six Gabor ﬁlters are

used to obtain the competitive code features with the

orientation set as θ

= π j/6, j = 0,1,...,5. u and σ

are ﬁxed to 0.0916 and 5.6179 respectively; All the

code features mentioned here are down-sampled by

ratio of 4:1. For obtaining the matching distance, the

code feature matrix needs to be shifted by rows and

columns. Here the shifting ranges are set to 0, [-1,1]

and [-2,2]. Time cost for identifying per test sam-

ple is recorded by Matlab 7.5 from a personal com-

puter with an Intel Pentium(R) Dual-Core Processor

(E5200@2.50GHz) and 2GB RAM conﬁgured with

Microsoft Windows XP.

When the system works in the identiﬁcation mode,

ﬁve samples from session one are randomly collected

to construct the training set, and all the samples from

session two are used for testing. The average identi-

ﬁcation rates (%) of ten-run executions with different

methods are listed in Table 1. Besides, the time cost

of identifying one test sample is also compared un-

der the current data set. When the system operates in

the veriﬁcation mode, all the 7752 samples are used.

Each palmprint template is matched with all the other

ones to evaluate the receiver operating characteristics

(ROC). A genuine matching is deﬁned as the match-

ing between the features from the same palm, and oth-

erwise the matching is counted as an impostor. The

total number of matches is 30 042 876, of which the

number of genuine is 74 068 and the number of im-

postor is 29 968 808.

Table 1 shows the comparisons of the average

identiﬁcation accuracy (%), time cost (s) for identi-

fying per test sample, and the veriﬁcation EER (%).

As can be seen, the recognition performance of cod-

ing based methods depends badly on the transla-

tion extent of code feature plane. Furthermore, the

matching speed dramatically drops as the translation

range increases. The identiﬁcation rate of CompCode

achieves 99.11% at the time cost of 2.398 seconds,

which is not desirable when the system is operated

in the identiﬁcation mode. DFT PalmCode outper-

forms PalmCode signiﬁcantly in terms of identiﬁca-

tion performance and veriﬁcation accuracy because of

two reasons. First, DFT PalmCode fuses more phase

information from two Gabor ﬁlters with different ori-

entations, instead of one Gabor ﬁltering as Palm-

Code exploits, which makes the discriminative ability

more powerful. Second, using the frequency-domain

representation of code features makes it capable of

undergoing training process by some learning algo-

rithms, which improves the recognition performance

and also reduces the time cost of dissimilarity calcu-

lation. By comparing the identiﬁcation performances

of CompCode and DFT CompCode, it can be seen

that the identiﬁcation accuracy does not improve ev-

idently while the matching speed drops dramatically

after executing DFT transform and (2D)

PCA on the

code features. This may be because the processes of

DFT transform and (2D)

PCA do not extract more

discriminative information from the image, but their

resulting real-valued feature templates cost less than

the logical-valued code feature planes. From Table 1

it can be also seen that lower EER is achieved by op-

erating the DFT and (2D)

PCA transform on the code

features. Since the matching score measurements for

coding based methods and the proposed feature repre-

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

(a)

(b)

Figure 7: Comparison of the veriﬁcation accuracies (%)

(FAR versus FRR). For the coding based methods, the code

template is shifted horizontally and vertically for multiple

times within a range [-2, 2].

sentation framework are totally different, we plot their

ROC curves for comparisons further. From Fig. 7(a)

and (b), we can see that DFT CompCode does not

improve the veriﬁcation accuracy of CompCode ab-

solutely though it does achieve lower EER. However,

DFT PalmCode does perform much better than Palm-

Code.

5.3 Comparisons with other Methods

Gabor ﬁltering based methods have been widely in-

vestigated in various image representation ﬁeld. Be-

sides coding based methods, there are two other

classes: First, Gabor coefﬁcients based statistical

learning methods, which make the statistical analy-

sis of the Gabor magnitude coefﬁcients (R. Chu et al.,

2007; X. Pan et al., 2008; X. Pan et al., 2009; M. Mu

et al., 2011); Second, local Gabor pattern histogram

methods (W. Zhang et al., 2009; B. Zhang et al., 2007;

S. Xie et al., 2010; W.C. Zhang et al., 2005), which

fuse the Gabor coefﬁcients with different local pat-

tern operators to achieve histogram sequences as de-

scriptors. In this section, comparisons with some of

them are evaluated on the palmprint recognition per-

formance. Note that for principal component analy-

sis (PCA) and linear discriminant analysis (LDA) in-

volved in these methods, we take the sum of the ﬁrst

largest eigenvalues that determine 90% of the total

sum of the eigenvalues for fair comparisons.

GM (2D)

PCA (X. Pan et al., 2008), as a repre-

sentative method of Gabor coefﬁcients based statisti-

Phase

quadrant

demodulation

coding

LXP

Concatenating

Real code

Imaginary

code

Gabor

part

Re_LGPP

Im_LGPP

Re_GGPP

Im_GGPP

HGPP

LBP

Gabor

phase

LGBP_Phase

Quantizing LBP LGXP

Figure 8: An illustration of various local pattern operators

compared in this paper. (LBP - local binary pattern, LXP -

local XOR pattern).

cal learning methods, is compared in this group of ex-

periments. A bank of ﬁve-scale and eight-directional

ﬁlters (in total 40 ﬁlters) is ﬁrstly used to derive a

Gabor feature space of high dimensionality and then

two steps of 2DPCA as described in section 4 are car-

ried out to reduce the dimension. The Gabor param-

eters are set as follows: θ

= π j/8, j = 0,1,...,7,

= 0.2592/

√

, v = 0, 1, . . . ,4, σ = 5.6179. For

GM (2D)

PCA method, the Gabor coefﬁcients are

down-sampled by the ratio of 4:1. Otherwise, the Ga-

bor feature dimensionality will be too high to be pro-

cessed further in our experimental system.

By encoding Gabor magnitude and phase co-

efﬁcients via local binary pattern (LBP) operator,

LGBP Mag (W.C. Zhang et al., 2005) and LGBP Pha

(W. Zhang et al., 2009) have been proposed respec-

tively. By using the local XOR pattern (LXP) oper-

ator to encode the real and imaginary parts of Gabor

complex response, local Gabor phase pattern (LGPP)

has been proposed (B. Zhang et al., 2007), denoted

by Re LGPP and Im LGPP respectively. Different

from LGPP, GGPP has been proposed to represent

orientation patterns, which forms one eight-bit binary

string to represent each pixel by concatenating the real

or imaginary quadrant-bit codes of different orienta-

tions for a given frequency (denoted by Re GGPP and

Im GGPP). In addition, Xie (S. Xie et al., 2010) pro-

posed local Gabor XOR pattern (LGXP) by quantiz-

ing the Gabor phase in each local region ﬁrstly, and

then encoding the quantized phases by LXP operator.

Figure 8 gives an illustration of different local Gabor

pattern operators. For fair comparisons, the Gabor pa-

rameters used in all the mentioned local Gabor pattern

methods are the same as those set in GM (2D)

PCA

method.

We assume that the system works in the identiﬁ-

cation mode, samples from session one are used for

training, and all the samples from session two for

test. Different training sets are constructed respec-

tively with the sample numbers of each palmprint

class ranging from two to ﬁve. The training samples

are randomly selected and ten-run executions are av-

FourierSpectralofPalmCodeasDescriptorforPalmprintRecognition

Table 2: Comparisons of the average identiﬁcation accuracy (%) and feature length among different methods. For evaluating

the identiﬁcation performance, the training sample number per class ranges from two to ﬁve. The average identiﬁcation rate

is calculated over ten-run executions with various training and testing sets.

Method

Identiﬁcation rate (%) Gabor

number

Feature length

before reduction

Final

feature length5 4 3 2

DFT PalmCode 99.74 99.67 99.55 99.24 2 7 936 2 878

GM (2D)

PCA 81.42 80.34 77.65 74.84 40 40 960 13 680

Re LGPP 97.10 96.87 96.20 94.18 40 40 960 1 540

Im LGPP 97.22 97.10 96.56 95.48 40 40 960 1 540

HGPP 98.40 98.11 97.93 97.82 40 92 160 1 540

LGBP Phase 97.75 97.57 97.08 95.92 40 40 960 1 540

LGXP 98.24 97.97 97.42 97.15 40 40 960 1 540

eraged to evaluate the identiﬁcation rate (%). Besides,

the resulting feature size is also investigated for eval-

uating the computational complexity.

Table 2 lists the comparisons of palmprint iden-

tiﬁcation performance. As it shows, the proposed

method achieves the highest identiﬁcation rate by

a large margin. Even under the condition of

less training samples, the identiﬁcation accuracy of

DFT PalmCode method reaches up to 99.24%. Dif-

ferent from the other Gabor based methods, which ex-

tract multi-scale and multi-directional features using

40 Gabor ﬁlters, DFT PalmCode uses only two Gabor

ﬁlters. Since the Gabor ﬁltering is time-costing, less

used ﬁlters leads to the advantages of higher execu-

tion speed and lower storage requirement. Besides, all

the local Gabor pattern methods generate a histogram

feature space with high dimensionality. As the num-

ber of resulting sub-blocks increases, the histogram

dimension will extend in multiple. Given the sub-

block number be N, and the image class number C,

then the ﬁnal feature vector length will be N ×(C −1).

As Table 2 shows, N ×(C −1) = 4×(386−1) = 1540

is the ﬁnal feature size of the mentioned local Gabor

pattern methods in this group of experiments, while

the ﬁnal feature length of DFT PalmCode is 2878,

which is larger but still comparable.

6 CONCLUSIONS

In this paper, we have presented a new feature ex-

traction method for palmprint recognition. The Gabor

phase information is ﬁrstly encoded into binary code

features, which are then transformed into Fourier

spectral (denoted by DFT PalmCode) as palmprint

descriptor. The spectral features are further processed

by the horizontal and vertical 2DPCA transform for

palmprint recognition. Experimental results demon-

strate its high efﬁciency under both veriﬁcation and

identiﬁcation system mode.

The main contributions and conclusions of this pa-

per are as follows: (1) The proposed method is related

to the original PalmCode approach, but they are quite

different. First, the real-valued spectral features are

achieved for palmprint recognition instead of the bi-

nary code matrices, which not only keep the high dis-

criminative ability of PalmCode features, but also can

be processed by some learning algorithms to further

improve the recognition performance. Second, due

to the translational invariance property of Fourier and

(2D)

PCA transform, DFT PalmCode method over-

comes the weakness of pixel-to-pixel matching strat-

egy, which is widely applied by the coding based

methods as well as PalmCode. Third, under the pro-

posed feature extraction framework, more than one

Gabor ﬁlters can be utilized to fuse phase informa-

tion, which improves the recognition accuracy evi-

dently compared with PalmCode. (2) The contrast

experimental results between another coding method

and our method demonstrate that the proposed feature

extraction framework can be used on other code-plane

based features for increasing identiﬁcation speed as

well as PalmCode. (3) Compared with other state-

of-the-art Gabor phase based methods which gener-

ally use 40 Gabor ﬁlters, the proposed method greatly

outperforms in terms of recognition accuracy, compu-

tational complexity and storage requirement by using

only two Gabor ﬁlters.

ACKNOWLEDGEMENTS

This work is supported partly by the National Grand

Fundamental Research 973 Program of China un-

der Grant No. 2004CB318005, and the Fundamen-

tal Research Funds for the Central Universities (Grant

No. KKJB11034536).

REFERENCES

Kong, A. and Zhang, D. (Aug. 23-26, 2004). Competitive

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

coding scheme for palmprint veriﬁcation. In Proc.

17th Int. Conf. Pattern Recognition (ICPR’04), pages

520–523, Cambridge, UK.

Kong, A. and Zhang, D. and Kamel, M. (2006). Palm-

print identication using feature-level fusion,. Pattern

Recognition, 39(3):478–487.

Zhang, B. and Shan, S. and Chen, X. and Gao, W. (2007).

Histogram of gabor phase patterns (hgpp): A novel

object representation approach for face recognition.

IEEE Trans. image process., 16:57–68.

Han, D. and Guo, Z. and Zhang, D. (Oct.26-29, 2008). Mul-

tispectral palmprint recognition using wavelet-based

image fusion. In Proc. Int. Conf. Signal Processing,

pages 2074–2077, Beijing,China.

Zhang, D. and Kong, A. and You, J. and Wong, M. (2003).

Online palmprint identiﬁcation. IEEE Trans. Pattern.

Anal. Mach. Intell., 25(9):1041–1050.

Zhang, D. and You, X. and Wang, P. and Yanushkevich,

S. N. and Tang, Y. (2009). Facial biometrics using

nontensor product wavelet and 2d discriminant tech-

niques. Int. J. Patt. Recog. Art. Intel., 23(3):521–543.

Daugman, J. (1993). High conﬁdence visual recognition

of persons by a test of statistical independence. IEEE

Trans. Pattern Anal. Mach. Intell., 15(11):1148–1161.

Yue, F. and Zuo, W. and et al.. (2011). Fast palmprint iden-

tiﬁcation with multiple templates per subject. Pattern

Recognition Letters, 32:1108–1118.

Yue, F. and Zuo, W. and Zhang, D. and Wang, K. (2009).

Orientation selection using modiﬁed fcm for competi-

tive code-based palmprint recognition. Pattern Recog-

nition, 42(11):2841–2849.

Yue, F. and Zuo, W. and Wang, K. and Zhang, D. (Jun.

24-26, 2008). A performance evaluation of ﬁlter de-

sign and coding schemes for palmprint recognition.

In Proc. 19th Int. Conf. Pattern Recognition (ICPR),

pages 1–4.

Beck, J. and Sutter, A. and Ivry, R. (1987). Spatial fre-

quency channels and perceptual grouping in texture

segmentation. Computer Vision, Graphics, Image

Processing, 37(1987):299–325.

Yang, J. and Liu, C. (2007). Horizontal and vertical 2dpca-

based discriminant analysis for face veriﬁcation on a

large-scale database. IEEE Transactions on Informa-

tion Forensics and Security, 2(4):781–792.

Mu, M. and Ruan, Q. (2011). Mean and standard deviation

as features for palmprint recognition based on gabor

ﬁlters. Int. J. Patt. Recog. Art. Intel., 25(4):491–512.

Chu, R. and Lei, Z. and Han, Y. and He, R. and Li, S. Z.

(Nov.18-22, 2007). Learning gabor magnitude fea-

tures for palmprint recognition. In Proc. 8th Asian

Conf. Computer Vision (ACCV2007), pages 22–31,

Tokyo, Japan.

Rowe, R. K. and Uludag, U. and Demirkus, M. and

Parthasaradhi, S. and Jain, A. K. (Sept.11-13, 2007).

A multispectral whole-hand biometric authentication

system. In Proc. Biometrics Symposium, pages 1–6,

Baltimore, MD.

Pankanti, S. and Bolle, R. M. and Jain, A. K. (2000). Bio-

metrics: The future of identiﬁcation. IEEE Computer,

33(2):46–49.

Xie, S. and Shan, S. and Chen, X. and Chen, J. (2010).

Fusing local patterns of gabor magnitude and phase

for face recognition. IEEE Trans. Image Process.,

19(5):1349–1361.

Jia, W. and Huang, D. and Tao, D. and Zhang, D. (12-15

Oct. 2008). Palmprint identiﬁcation based on direc-

tional representation. In Proc. IEEE Int. Conf. Sys-

tems, Man and Cybernetics (SMC2008), pages 1562–

1567, Singapore.

Jia, W. and Huang, D. S. and Zhang, D. (2008). Palmprint

veriﬁcation based on robust line orientation code. Pat-

tern Recognition, 41(5):1521–1530.

Shu, W. and Zhang, D. (1998a). Automated personal

identiﬁcation by palmprint. Optical Engineering,

37(8):2659–2362.

Shu, W. and Zhang, D. (Aug.16-20,1998b). Palmprint veri-

ﬁcation: An implementation of biometric technology.

In Proc. 14th Int. Conf. Pattern Recogn., pages 219–

221, Brisbane,Australia.

Zhang, W. and Shan, S. and Chen, X. and Gao, W. (2009).

Are gabor phases really useless for face recognition?

Pattern Analysis & Applications, 12(3):301–307.

Zuo, W. and Lin, Z. and Guo, Z. and Zhang, D. (Jun.13-

18, 2010). The multiscale competitive code via sparse

representation for palmprint veriﬁcation. In Proc.

23rd IEEE Int. Conf. Computer Vision and Pattern

Recognition (CVPR), pages 2265–2272, San Fran-

cisco, CA, USA.

Zhang, W. C. and Shan, S. G. and Gao, W. and Zhang,

H. M. (Oct.17-21, 2005). Local gabor binary pattern

histogram sequence (lgbphs): A novel non-statistical

model for face representation and recognition. In

Proc. 10th Int. Conf. Computer Vision, volume 1,

pages 786–791, Beijing, China.

Kong, W. K. and Zhang, D. and Li, W. (2003). Palm-

print feature extraction using 2-d gabor ﬁlters. Pattern

Recognition, 36(10):2339–2347.

Pan, X. and Ruan, Q. (2008). Palmprint recognition using

gabor feature-based (2d)

pca. Neurocomputing Let-

ters, 71(13-15):3032–3036.

Pan, X. and Ruan, Q. (2009). Palmprint recognition using

gabor-based local ivariant features. Neurocomputing

Letters, 72(7-9):2040–2045.

Wu, X. Q. and Wang, K. Q. and Zhang, D. (2005). Palm-

print authentication based on orientation code match-

ing. in: AVBPA 2005, Lecture Notes in Computer Sci-

ence, 3546:555–562.

Wu, X. Q. and Wang, K. Q. and Zhang, D. (Nov.3-6, 2006).

Palmprint texture analysis using derivative of gaussian

ﬁlters. In in: Proc. IEEE Int. Conf. Computational In-

telligence and Security, pages 751–754, Guangzhou,

China.

Hao, Y. and Sun, Z. and Tan, T. (Nov.18-22, 2007). Com-

parative studies on multispectral palm image fusion

for biometrics. In Proc. 8th Asian Conference on

Computer Vision (ACCV), pages 12–21, Tokyo, Japan.

Guo, Z. and Zhang, L. and Zhang, D. (Aug. 23-26, 2010).

Feature band selection for multispectral palmprint

recognition. In Proc. 20th Int. Conf. Pattern Recog-

nition (ICPR), pages 1136-1139.

FourierSpectralofPalmCodeasDescriptorforPalmprintRecognition

Guo, Z. and Zhang, D. and Zhang, L. and Zuo, W.

(2009). Palmprint veriﬁcation using binary orienta-

tion co-occurrence vector. Pattern Recognition Let-

ters, 30:1219–1227.

Guo, Z. H. and Lu, G. M. (July 11-14, 2010). Palmprint

recognition using gabor magnitude code. In Proc.

IEEE Int. Conf. Machine Learning and Cybernetics

(ICMLC), volume 2, pages 796–801, Qingdao, China.

Zhang, D. (2004). Palmprint Authentication. Kluwer Aca-

demic Publication, Dordrecht.

Sun, Z. N. and Tan, T. N. and Wang, Y. H. and Li, S. Z.

(Jun.25, 2005). Ordinal palmprint representation for

personal identiﬁcation. In Proc. IEEE Int. Conf. Com-

puter Vision and Pattern Recognition, San Diego, CA,

USA.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods