Palm Vein Recognition based on NonsubSampled Contourlet

Transform Features

Amira Oueslati

, Nadia Feddaoui

and Kamel Hamrouni

LR-SITI Laboratory, National Engineering School of Tunis, University ELMAanar, BP 37 Belvedere 1002, Tunis, Tunisia

ISD, University Manouba, 2010 Manouba, Tunisia

Keywords: Palm Veins Recognition, NonsubSampled Contourlet Transform (NSCT), ROI Extraction, Feature

Extraction, Feature Matching.

Abstract: This paper presents a novel approach for person recognition by palm vein texture image based on Nonsub-

Sampled Contourlet Transform (NSCT). Our approach consists of four steps. First, we reduce noise and en-

hance contrast in order to produce a better quality of palm vein image then we localize the texture in the

ROI. Next, the texture of enhanced image is analyzed by NSCT and obtained features witch are encoded to

generate a signature of 676 bytes. Finally, we compute hamming distance in comparison to take decision.

The experiments are performed on CASIA Multi-Spectral Palm print Image database. The method evalua-

tion is completed in both verification and identification scenarios and experimental results are compared

with other methods. Experiments results prove the effectiveness and the robustness of NSCT method to ex-

tract discriminative features of palm veins texture.

1 INTRODUCTION

Biometrics refers to all techniques to identify a per-

son based on its intrinsic characteristics that must be

unique and measurable. These features can be physi-

cal, biological or behavioral. The most known are

fingerprints, voice prints, iris, retina and hand. They

offer an irrefutable proof to distinguish one person

from another.

Therefore, palm veins recognition represents a

new generation of integrated access control system

based on the safest biometric technology today.

Palms have a large and rich blood vein patterns. It’s

unique to each person, stable during the person's

lifetime and a palm vein image is easily captured by

near-infrared rays, without contact and without a

trace. This makes this technology non-invasive,

hygienic and widely acceptable to users.

In this paper, to obtain a high-performance of

palm veins recognition system, we have applied the

Non sub sampled Contourlet Transform (NSCT)

method which is a shift-invariant, multi-scale, and

multi-directional transform. It can capture significant

veins features along all directions.

The rest of this paper is structured as follows. In

Section 1, existing methods in literature are briefly

reviewed, in Section 2 the proposed palm veins

recognition method using the NSCT is presented, in

Section 3 experimental results of the proposed

method are given and discussed. Finally, in Section 4,

conclusions are drawn.

2 STATE OF ART

Extensive work has been made on person recognition

using palm-vein technology based on many and dif-

ferent type of filtering. For example, In the work

presented by (Pan and Kang, 2011) the image is pre-

processed by histogram equalization, then three algo-

rithms (Scale Invariant Feature Transform, Speeded-

Up Robust Features and Affine-SIFT) were used to

extract local features, and finally the matching results

were obtained by computing the Euclidean distance.

Palm vein feature extraction from near infrared

images is proposed by (Sadeghi and Drygajlo, 2011)

an approach based on local texture patterns is

proposed. The operators and histograms of multi-

scale Local Binary Patterns are investigated to

identify statistical descriptors for palm vein patterns

and novel higher-order local pattern descriptors based

on Local Derivative Pattern histograms are then

Oueslati, A., Feddaoui, N. and Hamrouni, K.

Palm Vein Recognition based on NonsubSampled Contourlet Transform Features.

DOI: 10.5220/0005780902490254

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 249-254

ISBN: 978-989-758-175-5

249

investigated for palm vein description. In the work of

(Han and Lee, 2013), they proposed an adaptive

Gabor filter method to encode the palm vein features

in bit string representation. The bit string

representation, called VeinCode, offers speedy

template matching and enables more effective

template storage. The similarity of two VeinCodes is

measured by Hamming distance.

In this paper, a new approach is proposed for

palm vein recognition with a high performance based

on NonsubSampled Contourlet Transform.

3 PROPOSED METHOD

In biometrics, Palm vein technology works by identi-

fying the vein in an individual's palm. When a per-

sonal's hand is held over a scanner, a near-infrared

light locate the veins.

The red blood cells in the palm veins absorb the

rays and show up on the map as black lines, whereas

the hand structure shows up as white. This vein

pattern is then verified to authenticate the individual.

Figure 1: The process of the Palm veins recognition.

Figure 1 shows the general process of the

recognition model using Palm veins biometrics. This

palm vein recognition system consists of four steps:

palm veins region localization, preprocessing, feature

extraction, and matching.

3.1 Preprocessing

First, to segment the input image from the back-

ground we have applied a threshold. The smaller

objects due to the noise are removed through con-

nected components labeling. To normalize the con-

tour of the hand image a morphological operator

closing is carried out with a square structuring ele-

ment.

3.2 ROI Extraction

This process has many important aims; first, it serves

to remove the rotation, translation and scale (RST)

variations of palm vein images. Second, it allows

extracting the most informative area in the palm vein

image.

In the proposed palm vein recognition method, we

apply the method of (Feng et al. 2011). The four steps

to obtaining the square area ROI are described as

follows:

1) Line up the key point C

and the key point C

get the X-axis of image and then make a line

through the key point C

, perpendicular to the Y-

axis, and the intersection is d

2) Calculate the mean δ distance between the points

, C

and make:







(1)

3) Locate the square a

, a

and C

are

parallel, and make:





σ













(2)

4) Get the ROI from the original image, and

transform it into a grey image of fit size.

Figure 2: Region of interest (ROI) of palm vein image.

3.3 Palm Veins Feature Extraction

We analyze the texture to detect the most distinctive

characteristics using NSCT to extract texture features.

3.3.1 Texture Description based on NSCT

NSCT is composed of two-channel nonsubsampled

filter bank (NSFB). One is the Nonsubsampled direc-

tional filter banks (NSDFB) that provide the direc-

tionality and the other one is the Nonsubsampled

pyramid (NSP) that ensures the multi-scale property

(

Gonzalez et al., 2010)

. The structure of the NSCT (Fig.

3) ensures the shift-invariant property. Figure 3(a)

shows the structure of the two-channel NSFB, and

Figure 3(b) shows a 2-D frequency domain split into

a number of subbands. The NSP consists on a high-

pass (HP) subband and a low-pass (LP) subband, and

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

250

the NSDFB decomposes the HP subband into a num-

ber of directional subbands.

Consequently, the NSCT is effective in

representing well the detailed characteristics of the

strong palm vein texture along the radial and angular

directions. (

Zhou et al., 2012); (Tang et al., 2007); (Yang

et al., 2007).

(a) NSFB that implements the

NSCT

(b)2-D frequency partitioning

domain into a number of

subbands

Figure 3: NSCT structure (Gonzalez et al., 2010).

3.3.2 Palm Vein Features Extraction using

the NSCT

The binary vein image Fig 4(c) is used as the input of

the NSCT. After filtering by the NSCT, all of the

NSCT coefficients are used to define the palm vein

features.

The NSCT coefficients in all the subbands can be

defined as:

W









1jJ

1id





(3)

(J) is the total number of scale decomposition; (dj) is

the number of directions at j-th scale. W

Represents

the low-frequency coefficients and W

represent the

mid/high-frequency coefficients in the i-th directional

subband of the j-th scale level.

To capture strong distinct directional

characteristics of the palm vein, the coefficients in the

HP subbands





at scale j are used as the palm vein

features. Little directional information is contained

because the pyramid subband is an approximation

and contains LP information. Therefore, the LP

coefficients are W

excluded in the palm vein feature

extraction process. (

Gonzalez et al. 2010)

Consequently, only the coefficients

W





in mid-

and HP subbands are used as palm vein features.The

results of the decompositions is shown in Fig4.

Then, a binary feature vector must be created, the

signs of the NSCT coefficients W





in each subband

are used to generate the binary code (BC):

Figure 4: the Directional subband images at three different

scale levels. (a) Scale 1, two directions (d1=2), (b) Scale 2

Four directions (d2=4), (c) Scale 3 Eight directions (d3=8).







x,y





1, i





x,y0

0, else

;1id



(4)

(x,y) represent the coordinates in each

nonsubsampled subband image





. It is known that

the resulting (BC) contains sign information in each

NSCT subband. All the directional characteristics in

both multi-directions and multi-scale are considered

in the generated (BC). The final binary palm vein

vector is expressed as:



x,y





bc





x,y



⋯bc









x,y



⋮bc







x,y



⋯bc









x,y





⋮⋯⋮bc







x,y



⋯bc









x,y







(5)

3.4 Palm Vein Feature Matching

In order to match two palm vein feature vectors, the

Hamming distance (HD) is used. The HD measure

between two palm vein feature vectors V

and V

can

be defined as: (

Daugman, 1993)

HD

V



x,

⊕V



x,





;

(6)

Where (x, y) represents the pixel coordinates in the

X×Y subband image. The HD measure between two

palm vein vectors calculates how many bits are

different, if the value of the HD is closer to ‘0’, it

means that the two palm vein vectors come from the

same subject, and vice versa.

Palm Vein Recognition based on NonsubSampled Contourlet Transform Features

251

4 EXPERIMENTAL RESULTS

In this section, we will introduce the performance

measurement of our algorithm in verification and

identification scenarios. Then we will try to compare

our results with results of well-known veins recogni-

tion methods found in the literature. Tests are carried

out on left and right hands images of CASIA Multi-

Spectral Palm print database (

CASIA Multispectral

Palmprint Database).

The CASIA database contains large RST

variations since it was acquired using a non-contact

sensor, it has 200 identities, 6 samples per identity

witch give us 1200 palm veins.

Verification experiments and identification

experiments are detailed in the Section below.

4.1 Verification Mode

We perform verification experiments by reporting the

inter-class and intra-class curves, Receiver Operating

Characteristic (ROC) curves, the decidability (MD),

the degree of freedom (DOF) and the equal error rate

(EER).

In verification mode, all images in the database

are matched to all other images which commonly

named as “all versus all” in the literature.

After the calculation of the hamming distances of

all templates comparisons, which is a fractional

measure of dissimilarity; 0 would represent a perfect

match, the Intra-Class and the Inter-Class distribution

are found. The Inter-Class distribution of the

hamming distances is generated by comparing

between templates of the different persons. Intra-

Class distribution is generated by comparing between

different templates of the same sample. The number

of operations is 359400, 3100 of genuines and

716 300 of imposters.

To test the separability of the veins recognition

system, the decidability index MD proposed by

(Daugman, 1993) is used. The mean and the standard

deviation of intra-class and inter-class distributions

are calculated in order to calculate the decidability, If

their two mean values are µ1 and µ2, and their two

standard deviations are σ1 and σ2, then ‘MD’ is

defined as:

MD



μ









σ



(7)

We introduce also the uniqueness of the vein patterns

which means that there is an independent variation in

the vein details (Sun et al., 2005); we can determine

the vein uniqueness by examining the Inter-Class

distribution and calculating the degree of freedom

‘DOF’ defined as:

DOF



1μ







(8)

To choose the best number of decomposition k, we

have varied k form 2 to 4 and for each stage we have

generated the intra class and inter class distributions

as shown in Fig. 5. The obtained decidabilities and

the degrees of freedom are given in Table 1.

Table 1: Decidability ‘MD’ and degree of freedom ‘DOF’.

Stage k MD DOF

k = 2

1.22 11.06

k = 3

3.46 72.12

k = 4

3.43 50.67

The measure of decidability achieved high value

on CASIA database from k equal to 3; it reflects the

perfect separation between the two distributions

which don’t overlap. So the less error is found which

allows for more accurate recognition. We will focus

our study on the number of decomposition stages

k=3.

Figure 5: Intra class and Inter class distributions.

The highest HD of the intra-class distribution is

0.21 and the smallest HD of the inter-class

distribution is 0.2. Thus, a decision criterion of 0.2

can perfectly separate the dual distribution. The error

rates FAR (False Accepted Rate) and FRR (False

rejected Rate) will be 0.2% at that threshold.

Clearly, the probability of not making a false

match for single one-to-one verification trials is

99.70%, so we can conclude that the probability of

making at least one false match when searching a

database of N unrelated patterns is 0.30%.

The veins recognition system error rates, FAR and

FRR, are dependent on the adjustable adopted

threshold. Our algorithm will then compute the HDs

that will be sorted in ascending order and vary the

threshold from 0 to 1 then determine in each case the

corresponding FAR and FRR which will be used to

construct the ROC curve shown in Fig. 6.

When we increase threshold value, the FAR will

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

252

increase and FRR will decrease. When FAR is equal

to FRR, this value is called ERR (Equal Error Rate).

Figure 6: Comparison of ROC curves with other tech-

niques.

The TAR (True Acceptance Rate) can be used as

an alternative to FRR while reporting the perfor-

mance of a biometric verification system.

The results show a 99.70% of TAR in our

proposed system.

4.2 Identification Mode

The methodology we have followed consists on ran-

domly selecting one of palm veins images per subject

to form the gallery and treats all the remaining imag-

es as probes. The training set is composed of 200

Right/200Left samples and the testing set is com-

posed of 400 Right/400Left.

We report cumulative Match Characteristic

“CMC” curve (fig. 7). This system achieved 99.80%

of rank-one identiﬁcation rate.

Figure 7: Comparison of CMC curves with other tech-

niques.

4.3 Comparison with State-of-the-Art

Methods

We compare the performance of the proposed method

with three state-of-the-art ones namely Wavelet

(Kong et al., 2004) Gabor Wavelet Filter (Sun et al.,

2005) and Derivative of Gaussian Filter (Wu et al.,

2006). Tests are carried out on CASIA database.

A direct comparison on ROC curves is shown in

Fig. 6 and the results of the Equal Error Rate (EER)

and the TAR are summarized in Table 2.

Table 2: EERs and TARs of the different techniques.

Method EER % TAR %

NSCT 0,2000 99,70

Derivative Of Gaussian Filter

2.8887 92.01

Gabor Wavelet Filter

0.8660 98.40

Wavelet

0.4999 98.66

We can observe that our method gives perfect

results. Indeed, the NSCT ROC curve coincides with

coordinate axis. When the threshold is low, FAR will

be 0 and FRR vary from 1 to 0.When the threshold is

high, FRR will be 0 and FAR vary from 0 to 1. When

the threshold is about 0.2, FAR and FRR will

intersect at the 0.2 position, so the EER rate will be

0.2000% and the TAR will be 99.70% witch

consistently outperforms all other techniques for

CASIA database. We observe clearly that our

proposed method achieves an important EER

reduction compared to the nearest competitor,

Wavelet method.

Figure 7 shows a comparison of the CMC curves

of these several techniques and Table 3 summarizes

the results of rank-1 identification rates and time

processing of different techniques.

Table 3: Comparison of different algorithms.

Method Rank 1 identification rate Time(s)

NSCT 99,80

0.0010

Derivative Of Gaussian

Filter

95.10 0.0900

Gabor Wavelet Filter

99.01 0.1519

Wavelet

99.63

0.1214

The NSCT code achieved 99.80% identification

rate for rank-1. It is fair to deduce that the proposed

method performs better for identification in

comparison to state-of-the art techniques. We make

also a comparison on the time of execution shown in

Table 3 which demonstrates that our approach is very

quick and perfect.

5 CONCLUSIONS

This paper proposes a palm vein recognition method,

in which shift-invariant, multi-scale, and multi-

Palm Vein Recognition based on NonsubSampled Contourlet Transform Features

253

directional NSCT coefficients are used as effective

palm vein features.

We apply a pre-processing of the image to

eliminate the noise and the unwanted points.

Next, all the NSCT coefficients in each

directional subband are used to extract palm vein

features. The created palm vein vector extracts

desirable characteristics of features in both multi-

scale and multi-directions. These features are encoded

to generate a signature of 676 bytes. Finally,

hamming distance is computed in comparison.

Experimental results show the effectiveness of the

proposed NSCT feature based method in verification

and identification modes. We obtain an excellent

results of 99,80% of rank one recognition rate and

0.2000% of EER.

Future research will focus on the improvement of

the execution time and the performance of the

proposed algorithm fusion with the multimodality of

palm and dorsal parts of hand.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the National

Laboratory of Pattern Recognition, Institute of Auto-

mation, Chinese Academy of Sciences for their sup-

ply of Casia Multi-Spectral Palm print image data-

base.

REFERENCES

M. Pan, W. Kang, 2011. Palm Vein Recognition Based on

Three Local Invariant Feature Extraction Algorithms

Biometric Recognition, Lecture Notes in Computer Sci-

ence, Vol. 7098.

M. M. sadeghi, L. Drygajlo, A. , Oct. 2011. Palm vein

recognition with Local Binary Patterns and Local De-

rivative Patterns, International Joint Conference

on Biometrics Compendium (IJCB), 11-13, Washing-

ton.

W.Y. Han, J. Ch. Lee, March 2013.Palm vein recognition

using adaptive Gabor filter, Expert Systems with Ap-

plications, Vol. 40.

YiFeng,Jingwen Li, YiFeng, Jingwen Li, Lei Huang, and

Changping Liu, 2011. “Real-time ROI Acquisition for

Unsupervised andTouch-less Palmprint,” World Acad-

emy of Science, Engineering and Technology 54.

R. C. Gonzalez and R. E. Woods, 2010. Digital image

processing, (Pearson Education Inc., Upper Saddle

River), NJ, USA, 3rd edn.

J.G. Daugman, Nov.1993. High confidence visual recogni-

tion of persons by a test of statistical independence,

IEEE Trans. Pattern Anal. Machine Intell, vol. 15, no.

11.

Y. Zhou and J. Wang, Nov. 2012. Image denoising based

on the symmetric normal inverse Gaussian model and

NSCT, IET Image Processing, Vol. 6, no. 8.

L. Tang, F. Zhao and Z. G. Zhao, Nov. 2007. The nonsub-

sampled contourlet transform for image fusion, Proc.

Int.Conf. Wavelet Analysis and Pattern Recognition,

Beijing, China.

B. Yang, S. T. Li and F. M. Sun, Aug. 2007. Image fusion

using nonsubsampled contourlet transform, Proc.

Fourth Int. Conf. Image and Graphics, SiChuan, China.

“CASIA Multispectral Palmprint Database.” [Online].

Available:http://www.cbsr.ia.ac.cn/MS Palmprint Data-

base.asp.

A.-K. Kong and D. Zhang, August 2004. Competitive cod-

ing scheme for palmprint verification, International

Conference on Pattern Recognition, Cambridge UK.

Z. Sun, T. Tan, Y. Wang, and S. Z. Li, June 2005. Ordinal

palmprint representation for personal identification,

Computer Vision and Pattern Recognition CVPR, San

Diego.

X. Wu, K. Wang, and D. Zhang, Nov. 2006. Palmprint

texture analysis using derivative of gaussian filters, In-

ternational Conference Computational Intelligence and

Security, New York.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

254