Iris Feature Extraction and Recognition Based on Wavelet-Based Contourlet Transform

(1)

Procedia

Engineering

www.elsevier.com/locate/procedia

2012 International Workshop on Information and Electronics Engineering (IWIEE)

Iris Feature Extraction and Recognition Based on

Wavelet-Based Contourlet Transform

Zhongliang Luo

*

School of Computer Science, Shaoguan University, Shaoguan 512005, China

Abstract

In view of the limitation of poor direction selectivity about 2-D wavelet transform and the problem of redundancy on contourlet transform, an iris texture feature extraction method based on wavelet-based contourlet transform (WBCT)for obtaining high quality features is proposed in the paper. Firstly, the preprocessed iris image is decomposed by WBCT, then calculating its energy, mean, standard deviation and Hu invariant moments of each subband of different scales and different directions, and taking them as the eigenvalues of iris image, finally, it tests on four iris image databases by using Euclidean distance. Experimental results show that the algorithm is simple and effective, and obtain better recognition performance.

Keywords: Iris recognition, Wavelet-based contourlet transform, Feature extraction; Texture

1. Introduce

Texture of iris image is similar to ridges, recesses, spots, filaments, mole, whirlpool, which contains rich frequency and space position information and can be seen as the iris feature information carrier. For better represent the iris texture feature and improve the recognition rate of iris recognition system, a better method of extracting iris space and frequency feature is needed.

Contourlet transform is a kind of double iterative filter structure that combined Laplacian pyramid(LP) with direction filter Bank (DFB)[1], developed on the basis of wavelet transform, which is another image representation method of multi-resolution, time-frequency local features and multi-direction. Due to

*

* Corresponding author.

E-mail address: luozl66@yahoo.com.cn.

Open access under CC BY-NC-ND license.

(2)

contourlet transform can extract the very important intrinsic geometric structure of image, it can describe the edge detail and texture of image with the best way. With constantly enrich and improve of contourlet theory and algorithms, it gradually reflect a better advantage in image denoising, enhancement and fusion[1-3].

However, contourlet transform has redundancy. So, a non-redundant multi-scale geometric transform is put forward in this paper, that is to say wavelet-based contourlet transform(WBCT), we can extract iris feature using WBCT. The method takes mean, standard deviation and energy of different scales and different direction coefficient matrix as iris feature value. Then, they are matched by using Euclidean distance. Feature dimension is less for these feature extracting method and they can speed classification. The experimental results verify effectiveness of WBCT during iris texture feature extraction.

2. Iris texture feature extraction

There is 4/3 redundancy for contourlet transform, it need apply Laplacian pyramid at its first stage of transform. So, it is not an ideal image sparse representation method. In 2004, Eslami and Radha proposed WBCT to solved the redundancy problem of contourlet transform, and it maintains a good nonlinear approximation of wavelet transform[4-6]. WBCT is a non-redundant completely reconstruction, for wavelet transform and DFB can achieve non-redundant perfect reconstruction. Features extracted using WBCT can better reflect the nature of iris, it is helpful to improve the stability and accuracy of iris recognition.

2.1 WBCT decomposition

It obtained subspaces after the image

I x y

( , )

is decomposed by wavelet transform, and

W

_{j HL}_, ,

, j LH

W

,

W

_{j HH}_, is the horizontal, vertical and diagonal details of image

I x y

( , )

at scale

j

.Their basis

functions is

ψ

_{j HL}_,

( )

n

,

ψ

_{j LH}_,

( )

n

,

ψ

_{j HH}_,

( )

n

respectively. It obtained the

k

-th direction subspaces after the three subspaces are adopted to

l

_j direction filter, they label l_,j _,

j HL k

W

, l_,j _, j LH k

W

and l_,j _, j HH k

W

（

0

≤ ≤

k

2

lj_）_{respectively, the relation between scale space}

j

V

and wavelet space

W

_j as follows

j j j

V

W

V

₋₁

=

⊕

(1) 1 2 , ₀ , , l j j l j i _k j i k

W

− =

= ⊕

，

i HL LH HH

=

,

(2) Where,

⊕

is space direct sum. After wavelet filter scale

j

and DFB direction filtering, basis functions of WBCT direction subspace are

2 , ,

( )

(

)

,

( )

j j j l l l j i k k k j i m z

n

g n S m

m

η

ψ

∈

=

∑

−

(3) Where, lj k

g

is the direction filter and lj k

S

is down-sampling matrix. After direction filtering with

l

_j level, it get subbands:

,

[ ]

,

, ,

( )

j j

l l

j k j i k

c

n

=

I

η

n

(4) Fig.1 indicates WBCT decomposition, where Fig. 1(a) show the WBCT decomposition structure, Fig.1 (b)shows the decomposition combined 2 level wavelet decomposition with 3 level DFB, direction

(3)

number of DFB is 8. As can be seen from the decomposition figure, WBCT is better express more sparse image than wavelet transform and contourlet transform, thus, it better extract the image texture feature, which will helpful to distinguish and identify iris.

(a) image decomposition structure by WBCT (b) decomposition coefficients of WBCT Fig. 1 WBCT decomposition

If the input image is

I x y

( , )

, then the WBCT decomposition results can be expressed as 1 ( , ) _J J _j j I x y a b = = +

∑

(5) Where,

a

_Jand

b

_jis the low-frequency subbands and high-frequency subbands respectively. 2.2 Feature extraction based on WBCT

As a non-redundant multi-scale geometric transformation, WBCT can better take advantage of the unique geometry of the image to express its profile and texture and provide essential information for iris recognition. It takes mean, standard deviation, energy and Hu invariant moments of subband as iris feature value respectively.

1) Mean and standard deviation

The mean

I

_ij and standard deviation

δ

_ij of each subband coefficients are defined as follows

1 1 2 1 1 1 ( , ) 1 [ ( , ) ] ij ij ij ij M N ij ij x y ij ij M N ij ij ij x y ij ij I I x y M N I x y I M N δ = = = = ⎧ = ⎪ ⋅ ⎪ ⎨ ⎪ ₌ ₋ ⎪ _⋅ ⎩

∑ ∑

(6)

Where,

I

_ijis the subbands of scale

i

and direction

j

,

M N

_ij

⋅

_ij represents the size of subband coefficients matrix

I

_ij.

2) The

l

₂ norm energy

The average energy of each subband coefficients matrix is defined in the following formula [7]

2 1 1 Mij _{| ( , ) |} ij ij x ij ij E I x y M N = = ⋅

∑

(7) 3) Hu moment invariants

In 1962, Hu demonstrated that moment invariants is not sensitive to shift, rotation and scale invariance, its has strong noise immunity and can effectively reflect the essential characteristics of

(4)

image[8]. So, Hu moment invariants is taken as effective features in the iris recognition.

For the image

I x y

( , )

, its size is

M N

×

, its

(

p q

+

)

order of the concerning moments

m

_pqand central moments

μ

_pq are defined as following

1 1 1 1

( , )

(

)(

) ( , )

M N p q pq x y M N pq x y

m

x y I x y dxdy

x x y y I x y dxdy

μ

= = = =

⎧

=

⎪

⎨

⎪

=

−

⎪⎩

∑∑

(8)

Where,

( , )

x y

is the area centroid coordinates,

x m m

=

₁₀ ₀₀

y m m

=

₀₁ ₀₀ ,

p q

,

=0,1,…。The normalized center distance of is defined as

00

pq pq

η

=

μ

Γ ₍

Γ =

[(

p q

+

) / 2] 1

+

_,

p q

+ =

2,3,

L

_{) (9)} Concerning moments and central moments can be used to describe the region shape, but they have not

invariant. So, Hu derived seven moment invariants set from the 2nd and 3rd order normalized central moments, they satisfy rotation, translation and scale invariant. The formula is as follows

1

η

20

η

02

Φ =

+

;

Φ =

₂

(

η

₂₀

−

η

₀₂

)

2

+

4

η

₁₁2 ; 2 2 3

(

η

30

3 )

η

12

(3

η

31

η

03

)

Φ =

−

+

−

;

Φ =

₄

(

η

₃₀

+

η

₁₂

)

2

+

(

η

₂₁

+

η

₀₃

)

2 2 2 5 30 12 30 12 30 12 21 03 2 2 21 03 21 03 30 12 21 03

(

3 )(

)[(

)

3(

) ]

(3

)(

)[3(

)

(

) ]

η

η η

η

η η

η

Φ =

−

+

−

+

−

+

−

+

2 2 6 20 02 30 12 21 03 11 30 12 21 03 2 2 21 03 30 12 21 03

(

)[(

)

(

) ] 4 (

)(

)

(

)[3(

)

(

) ]

η

η η

η

Φ =

−

+

−

+

−

+

2 2 7 12 30 30 12 30 12 21 03 2 2 21 03 21 03 03 12 21 03

(3

)(

)[(

)

3(

) ]

(3

)(

)[3(

)

(

) ]

η

η η

η

η η

η

Φ =

−

+

−

+

−

+

−

+

1 2 3 4 5 6 7

{ ,

,

}

Φ = Φ Φ Φ Φ Φ Φ Φ

(10)

In order to avoid the differences between the various features too large, feature vector

v

are needed to normalize, the normalized feature vector

v

′

is obtained according to the following formula

[

min( )] [max( ) min( )]

v

′ = −

v

−

v

(11) 3. Experimental results and analysis

Test images of algorithm simulation derive from four iris databases. It adopted the preprocessing method in reference [9] to preprocess iris image before test, feature extraction and recognition at the same experimental conditions, extracting features from each normalized iris image after image is decomposed with WBCT. The test number is 3、9、2 and 5 for CASIA Ver1.0, CASIA Ver2.0, MMU and self-acquire iris database, the others used to test. To verify the effectiveness of the algorithm, it compared iris feature extraction algorithm based on contourlet transform with WBCT.

Feature extracting by contourlet transform, its Laplacian pyramid selects "9-7" filters ,the directional filter uses "pkva" filters and decomposition layer is 5. However, iris feature extraction algorithm based on WBCT, it decomposed the preprocessed image with 2-layer wavelet decomposition, the wavelet basis is

(5)

‘db1’, which will get 48 subbands in all directions after the HH,HL,LH subband is decomposed by contourlet with 3 layer.

Taking energy, standard deviation and mean as feature, each high frequency and low frequency subbands features as the corresponding eigenvectors, the total number is 49. And calculating the moment invariants of each high-frequency subbands and low-frequency subband’s energy when Hu moment invariants are taken as the main feature, the total feature number is 337. it uses Euclidean distance to calculate the iris similarity for matching. As can be seen from Table 1, the recognition results show that the proposed method is better than the compared method, apart from the method of combination features, energy feature method is good, but features dimensions is high and the computation time increased for the features combination method.

Table 1. Recognition results of extracting iris feature based on contourlet transform / WBCT (%)

Iris image database energy mean standard deviation Hu moment

invariants energy+ standard deviation +Hu moment invariants CASIA Ver1.0 95.1/ 95.5 93.9/ 94.7 94.3/ 94.8 94.5/ 94.7 95.4/ 95.9

MMU 94.1/ 94.4 92.1/ 92.7 91.9/ 92.6 91.0/ 91.8 94.1/ 94.5

CASIA Ver2.0 95.0/ 95.3 94.8/ 95.1 94.9/ 95.1 94.9/ 95.1 95.1/ 95.5 Self-acquired 89.6/ 89.8 89.1/ 89.3 89.2/ 89.5 89.0/ 89.3 89.0/ 89.5

4. Conclusion

Considering the limitation of poor direction selectivity about wavelet transform and the problem of redundancy on contourlet, an iris texture feature extraction method that based on WBCT is adopted in this paper. Iris feature is obtained by calculating the mean, variance, energy and Hu moment invariants of each sub-band coefficient matrix. The feature extraction method can accurately express the feature of the iris texture with less information, at the same time, it has high recognition rate, the dimension of feature vectors is low, significantly reducing storage and computational complexity. However, the decision of the best decomposition layers and which subband contains the important information remains to be further studied.

References

[1] DoM N,Vetterli M. The contourlet transform: an efficient directional multiresolution image representation. IEEE Trans on Image Processing,2005,14(12):2091-2106

[2]Li X.M, Yan G.P Chen L. A New Method of Image Denoise Using Contourlet Transform. International Journal of Intelligent Information Technology Application, 2010, 3(1): 25-30

[3]Liu Z. Minimum distance texture classification of SAR images in contourlet do-main. Proceeding of International Conference on Computer Science Engineering and Software, Wuhan :China, 2008 ,1:834-837

[4]Miao Q.G., Wang B.S.. A novel image fusion method using WBCT and PCA. Chinese Optics Letters,2008,6(2): 104-107 [5]Gao X.B, Lu W, Li X.L.Wavelet-based contourlet in quality evaluation of digital images.Neurocomputing,2008,72:378-385 [6]Eslami R., Radha H. Wavelet-based contourlet transform and its application to image coding. Proceedings of International Conference on Image Processing. Singapore : IEEE Signal Processing Society , 2004. 3189- 3192.

[7] Kokare M., Biswas P K., Chatterji B N.. Texture image retrieval using rotated wavelet filters. Pattern Recognition Letters, 2007, 28(10): 1240-1249

[8] Hu M.K.. Visual pattern recognition by moment invariants. IEEE Trans. on Information Theory, 1962,8(2):179-187. [9] Luo Z.L, Lin T.S. Preprocessing of Iris Image for Identification Recognition . Journal of South China University of Technology,2009,37(5):43-47

Open access under CC BY-NC-ND license.