Multimodal Biometric: Iris and face Recognition
based on feature selection of Iris with GA and scores
level fusion with SVM
Yacine BOUZOUINA Latifa HAMAMI
Ecole Nationale Polytechnique, Ecole Nationale Polytechnique,
Alger, Algeria Alger, Algeria
[email protected] [email protected] Abstract— the purpose of this work is to develop a biometric
bimodal system for the verification of the person: Our bimodal Multimodal biometric systems perform better than single
system is going to combine the modality of the face with that of biometric systems where most of the Multimodal biometric
the iris. Firstly, the modality of the face constitutes one of the systems have used multiple modalities to overcome the
most natural ways to recognize a person. On the other hand, the limitations of the single biometric such as: noise, non-
modality of the iris is certainly one of the most used. We universality and spoof attacks[4].
introduce the recognition of the face by using a Discrete Cosine
Transform DCT, it transforms an image signal from a space It is necessary to normalize the region of the iris segmented
representation into a Frequency domain, and PCA transform, into a fixed size. It is performed by allocating to every pixel of
which treat the face image as a stochastic variable. The the iris in cartesian domain a correspondent in polar domain,
performance of the systems of recognition of the iris are this stage is in general fulfilled by the model used by Daugman
significantly improved by the precision of the segmentation with [6]..The next process is to enhance the normalized pattern by
Snake method. The fusion in feature level of two features applying histogram equalization in order to improve the global
extractors, called Gabor filter and Zernike moment through the contract of pattern. [7]
feature selection using a Genetic Algorithm in order to combine
the properties of global and local methods. Later we will study a The Zernike moment has been used in large scale of image
fusion at the scores level of the face and iris modalities after being processing works, and specifically used in biometric system,
normalized by using SVM classification method. The for example works of Chun-Wei Tan[17], he uses Log Gabor
performance of our system is tested using CASIA-IrisV3-Interval filter and ZMs for feature extraction of iris, the fusion of both
data base, the recognition rate achieved with this data base is methods are performed in score level, and Yarui Chen[15]
98.8. performs the ZM for face and Fingerprint recognition, in this
work the ZM is applying in feature extraction level with Local
Keywords— Biometrics, iris recognition, Zernike moment, Binary Pattern (LBP) and Log Gabor filter from both face and
PCA, DCT, SVM, Genetic Algorithm. fingerprint and the fusion through 2DPCA and NMF[15].
I. INTRODUCTION In this paper two modalities were chosen: iris and face, in
order to enhance the performance by combining information of
All biometric modalities have their own sensors, both modalities, we have used segmentation with Active
algorithms, strengths and weaknesses points, for example the contours by locating the initial stage of a recognition system of
fingerprints give a good precision of identification of the the iris as acquisition of the picture of the eye, and it is
individuals, even the footprints of the identical twins are necessary to perform a segmentation to isolate the iris from the
differential, besides it requires a big counting capacity and captured picture of the eye to keep only useful information,.
Footprint can be absolutely or partly distorted because of This process localizes the pupillary, and isolate limbic and
environmental and professional factors [1]. Although, the removes pupillary, specular reflections, eyelids, and the
available systems of recognition of the face gives a reasonable eyelashes [4]. A method proposed here will be demonstrated in
precision, they require some restrictions during the acquisition the next chapter [5].
of facial picture and the appearance of the face changes in the
course of time [2]. However, the iris is the annular region of There are several works in biometric system involved the
the eye delimited by pupil of inside and the sclera of outside, feature selection in order to enhance the performance by
among the advantages of this modality it does not change selecting the best set of features, Maryam Eskandari and Önsen
courtyard of time [3]. Toygar [2] uses the feature selection in face recognition, they
apply the feature selection to select best features extracted with
The systems of recognition of the iris commercialized PCA, LDA, spPCA, mPCA and LBP using Particle Swarm
require a collaboration of the users during the acquisition of Optimization (PSO) and Backtracking Search Algorithm [2]
picture, (without movement, background fixes, angle of
acquisition, conditions of illumination), this method is Feature selection is a corps of our work, we are going to
considerably more expensive [3]. use two methods of encoding iris, the global iris features are
extracted by employing 1D log-Gabor filter and the local iris
978-1-5386-0706-0/17/$31.00 ©2017 IEEE
features are extracted by Zernike moment. This combination defines internal pressures dependent on the contour itself.
allows us to link the advantages of global and local methods in However, external force defines pressures belonging to image
order to enhance the recognition rate, among the advantages of and independent of the contour
ZM feature extraction is the invariant rotation property that we The Active contour is represented by its energy as:
want Operates it in our shame, next we are going to select the
best features by using Genetic Algorithm.
Esnake=
We have used the recognition of the face which was
established by hybridization of global and local methods: PCA Eint(v(s)), Eint(v(s)) represent the external and internal energy,
and DCT Respectively [8]. The score level fusion gives the respectively and ={x(s),y(s)} is the curve, s is the length
best compromise between the wealth of information and the
easiness of realization, this step is established by using the of curve.
classification method, which is SVM [9].
Eint(v(s)) =
II. FACE RECOGNITION
The methods of recognition of the face can be separated in Eext(v(s)) = |∇(Gσ(x,y)*I(x,y))|2 (2.3)
two big families, global methods and local methods, global
methods are based on the techniques of statistical analysis. It is Where Gσ is Gaussian function of external energy with
not necessary to repeat some details points of the face, these standard derivation σ, ∇ is the gradient operator and *
methods use in most cases an analysis subspace of faces [10]. represent the convolution [12]
The local methods use information that they have on the
morphology of the face and based in general on points The internal force is represented to push the contour to form a
characteristics of the face. circle, which is the form of the edge of pupil, the flow chart
In this paper, we adopt the hybrid methods, this hybridation below detailed the deferent steps of segmentation. While
provides the opportunity to link the advantages of global and external force, it means to push the contour towards the
local methods by combining the detection of geometric features position of this edge in image by using gradients of pixels [13]
(PCA) with extraction of local characteristics of appearance
(DCT). This combination will improve the stability and the
performance of recognition during changes of pose. Eye Image (320*280)
DCT Algorithms transform an image signal from a space Pixels
representation into a frequency representation. The low-
frequency part is the most visually significant region, and the
higher frequencies represent the finer details. The coefficients
are classified in zigzag pattern such that they are in increasing Hough circle Transform
frequencies from the top corner [11].
PCA treat the face image as a stochastic variable, it based
on discrete K-L transform, big eigenvectors are in associate
with big eigenvalues, and they form the main components of Find the circle of the Find the circle of the
pupil Iris
the face called Eigen face [11].
III. IRIS RECOGNITION:
Initial snake of Pupil Initial snake of Iris
A. Iris segmentation:
Number of points of the snake (120) Number of points of the snake (120)
There are several works that offer a method of segmentation α= 0.05 β= 0.05 σ=15 α= 0.05 β= 0.05 σ=15
of iris region. Our method is affected by measuring an
integral-differential operator, the operator is based on
hypothesis that both borders (outside and internal) of the iris
Compute the Energy of Compute the Energy of
have a circular form [12]. Image Image
Active contour is one of the resent methods used for iris
segmentation. Ritter, Owens, Van Saarloos and Cooper were
the first to introduce a model of active contour in 1999, to
locate the edge of eyes [13]. The active contour loses its Find the contour of pupil Find the contour of Iris
shape, moves across image under the influence of two force
predefined, an internal force and another one external force, to
obtain an equilibrium, then the active contour contains group
of summits called the points of controls, and an Internal force
Segmented Iris
Figure 1. Diagram of segmentation of the eye by active contour.
There are two steps for the detection of the iris-pupil B. Iris Normalization
boundaries:
The normalization of the iris segmented into a fixed size is
Search the initial contour of the pupil: in this step, we used performed by allocating to every pixel of the iris in Cartesian
the Hough transform to find the coordinates of the pupil domain a correspondent in polar domain, this stage is in
circle, and then initialize the contour from these points [14]. general fulfilled by the model used by Daugman [15].
Since the eyelid is the dark part of segmented iris, applying
Find the true contour of the pupil using the active contour
low pass filter and binarizing image is a good choice for the
method and the initial contour of the previous step. The
mask as showed in fig.3.
pupil is the strongest contrast region, then it is natural that
external forces pushing the contour to this region. The
evolution of the contour is affected by weighting the
internal forces by low values with α= β = 0.05 = 0.5, σ = 15
and maximum number of iterations = 100[14]. Rectangle of the texture (r = 20, θ = 24)
Find the outer limit of the iris is more difficult than the
inner boundary because the intensity differences between
the irises, eyelids and eyelashes, usually this difference (a)S1003L05
Rectangle of the mask (r=20, θ=240)
obscure the outer limit of the iris.
Initial contour of the iris: we chose to initialize the contour Figure .3 image of the CASIA V3 base segmented by the method
outside the iris by the Hough transform, and then to search of the snake and Normalized
for the true contour of the iris: we use the initial contour
and applying the current contour to find the real outline of
the iris. C. Proposed feature encoding:
In the proposed scheme, feature level fusion is adopted.
We impose on the contour a low elasticity and low Therefore, this level of fusion contains richer information of
stiffness and we weighted an external force, so that outline biometric data, where the global features are extracted by
is not moving towards the pupil. α= β = 0.05 = 0.005, σ = employing 1D log-Gabor filter and the local features with
7, maximum number of iterations = 500 [14], (see figure 1) Zernike moment, this method allow to link the advantages of
results of different images segmented of CASIA V3 base is global and local methods by combining the detection of
shown in Figure 2. geometric characteristics with the extraction of local
Initial contour of the iris: we chose to initialize the contour characteristics of appearance. Both methods together is
outside the iris by the Hough transform, and then to search allowing to improve the stability of the performance of
for the true contour of the iris: we use the initial contour recognition during changes of pose, illumination and Iris
and applying the current contour to find the real outline of expressions. In order to overcome the high dimensional
the iris. We impose on the contour a low elasticity and low vectors resultant issue, we applied optimization method to
stiffness and we weighted an external force, so that outline select the optimized subset of methods, the optimization
is not moving towards the pupil. α= β = 0.05 = 0.005, σ = technique was adopted is GA as demonstrated in fig 6, in
7, maximum number of iterations = 500 [14], (see figure 1) order to minimizing both the number of features and
results of different images segmented of CASIA V3 base is classification error [16].
shown in Figure 2.
Figure 4. Image with Gabor filter code.
The magnitude of the Zernike’s moment, i.e. ∣ nm∣ is a
descriptor invariant to the image rotation [16].
The localized ZMs features are computed from the image
blocks A of size a×a extracted from the normalized iris image
I, moving at an interval of s pixels in both the horizontal and
vertical directions. The interval s is defined as half of the
block size, i .e. s = a/2, NB total number of extracted image
Blok {Ai}Nb, with order n and repletion m[17].
Figure 2. Various images of the CASIA V3 base
segmented by the method of snake.
by equal error rate (EER) and the algorithm stops at 200
evaluation.
Figure 5. Image of Zernike moment code. IV. MATCH SCORE LEVEL FUSION
In matching scores level fusion, we find different matchers, it
can be considered as the classification of the scores
Accept/Reject such as SVM, or combination of the scores to
D. Genetic Algorithm: provide a scalar score such as Sum Rule, Weighted Sum Rule,
and Product Rule. In this work, we are going to combine the
Like any other optimization algorithm, the Genetic scores of the modality of the face with that of the Iris using the
Algorithm (GA) is defined by chromosome encoding, fitness SVM classification method. In order to normalize the
function and Stopping condition, The GA evaluate the matching scores generated by Hamming distance, we use
population same as the human natural evaluation, the initial Tanh, which gives good results with noisy scores [2].
population is created randomly, it will be evaluated by the
fitness function, for a binary selection Algorithm the value ‘1’ D. Classification:
is affected to the particular feature selected, and ‘0’ to the
particular feature not selected in the chromosomal evaluation The SVM has been used as a new method for scores
[16]. classification [25], The goal of SVM classification is to find
the decision boundary with linear separation, the equation of
The GA begins by defining an array of variable values to be boundary is f(x)=wTx+b = 0 (3.1).
optimized (chromosome)
F(x) > 0 corresponding yi=1.
If length of Nvar Chromosome = [P1, P2, P3......, PNvar].
F(x) < 0 corresponding yi=-1.
Each chromosome has a fitness found by evaluating the fitness
function f , at P1, P2, P3......,PNvar : The SVM finds decision surface genuine/impostor by
performing the recognition pattern by certain vectors that Built
Fit = f (chromosome) = f (P1, P2, P3......,PNvar) (2.9) in training phase, called support vectors.
In our case, searching for the maximum elevation on a We will formulate the problem as: X= (xi) set of scores
biometric system requires a fitness function with input normalized by Tanh method, Y=(yi) {-1,1}, which represent
the class of each input vector.
parameters P1= L P2 = Z
The decision function is:
L= {F1,F2,……..,Fn} , Z={ F1,F2,……..,Fm }
f (x) = sgn( yi αi[K(x)]i,x) -b) (3.2)
Where Z and L are features extracted by computed ZMs and
LGs features respectively In our case, kernel function were adopted is K(x,y)=x.y.
The training of SVM started by associate 1 to the Iris of each
person that belong to training-client database and
-1 for the training-impostor database.
V. DATABASE AND EXPERIMENTAL RESULT:
For the Face test, we use the ORL database; it took 280
Figure 6. Iris feature selection using GA. images in the stage of study
(40 persons with 7 acquisitions per person). For every image
Algorithm: General Structure of optimizing Algorithm [17] of test [10].
Creation of Population For iris test, the database CASIA-IrisV3-Interval is adopted, it
selecting feature using took 4700 images in the stage of study (235 person with 20
Fit_Fun acquisitions per person) [24], for every image of test. The
Return (EER %) segmented Iris was normalized with size of 280×20, for
End employing ZMs, each image is divided to 224 regions. The
Stopping condition optimization of features is achieved by the concatenation both
of ZMs and LGs feature, the order and repletion are selected
In this study the initial population contain a random binary through the Fig.7 that shows the recognition accuracy with
digits with 20 Chromosome, fitness function are represented different orders and repetitions, the chosen pair of order and
repetition is (5,5), because this pair have the best accuracy
recognition as presented in fig.7. Our segmentation method
give good accuracy in comparison with that of Libor Masek as
demonstrated in Fig.9.
Figure 7. selection of ZMs
Figure 8. Convergence of GA Algorithm
The fig.8 illustrate that the EER of Ga approach is better than
EER of ZMs and LG approach, The EER is improved from
0.11 of ZM and 0.09 from LG to 0.06 from GA approach.
While, the EER from ORL data base of PCA and DCT method
is 0.24 as shown in Tab.1 The fusion of two modalities result
the EER=0.045 fig.11.
The set of experimental results are shown in Tab1 with and
without rotation of Iris. The accuracies without rotation of
face recognition, Iris with Gabor filter, Iris with ZMs and
SVM fusion of the Iris and face modalities is 85.83, 95, 90.1,
97.5% respectively at 0.01 FAR fig.10.
By introducing the rotation of iris, we notice enhancement of
The GAR in Iris recognition with Gabor filter, on the contrary
the GAR with ZM feature extraction is the same due to the
invariant rotation property as shown in Tab1.
Figure .9 ROC curves of different
Figure 11. Equal error rate. Figure .10 ROC curves of proposed
Tab 1. the equal error rates from various iris matching approaches CASIA-IrisV3-Interval
Proposed methods without rotation of Iris Proposed methods with rotation of Iris
Method EER% GAR at 0.01FAR% EER% GAR at 0.01FAR%
Log Gabor filter 0.09 90.1 0.09 93.4
Zernike Moument 0.11 95 0.11 95.2
Face recognition 0.24 85.83 0.24 85.83
Face + Iris 0.045 97.5 0.045 98.8
GA Feature Selection 0.06 0.06
Fig8 Block diagram of proposed scheme.
VI. CONCLUSION
[5] Ouabida, Elhoussaine Essadique,Abdelaziz Bouzid,
This paper has treated a multimodal biometric system, it is Abdenbi«Vander Lugt Correlator based active contours for iris
based on the fusion at features level of the modality of the Iris, segmentation and tracking»,Vol.71 ,pp383-395,2017.
our proposed optimization technique has reduced both the [6] Er. Vijay Dhir, Dr.(Ms) Maitryee Dutta «new method of
computing time and memory consumption using the Genetic iris recognition based on j.daugman’s principle» Second
Algorithm, only the feature that gives the best recognition rate International Conference on Emerging Trends in Engineering
are remaining. However, Segmentation that was used gives and Technology, ICETET-09: IEEE.
best results including good accuracy in comparison with that [7] Latha, L.Thangasamy, S.«A robust person authentication
of Libor Masek as demonstrated in Fig.9. Our extensive system based on score level fusion of left and right irises and
results have shown that it can improve the stability and the retinal features»,Vol.2,pp111-120,2010.
performance of recognition during changes of pose, [9]A. Ross and A. Jain, «Information fusion in biometrics»,
illumination and facial expression. Moreover, the fusion of Vol. 38, N° 12, pp. 2270–2285,2005.
both modality Iris and face increase again the accuracy of our [10] Nawaf Hazim, «Face Recognition using PCA-BPNN with
system. DCT Implemented on Face94 and Grimace Databases»,vol
142 No 6 pp 8-13, 2016. .
[11] Saboori, Arash Hosseini, S Abolfazl«Based on
REFERENCES Combination of DCT and PCA»,Vol.7,pp521-524,2014.
[12] Abdullah, Mohammed A M Dlay, Satnam S Woo,Wai L
[1] Heng Fui Liau « Feature selection for support vector Chambers, Jonathon A«Robust Iris Segmentation Method
machine-based face-iris multimodal biometric system » vol Based on a New Active Contour Force With a Noncircular
38, No 9 pp11105–11111, 2011 Normalization» pp1-14,2016.
[2] Maryam Eskandaria, Önsen Toygarb, «Selection of [13] Gang Zhang« Double-pupil location of face images »,vol
optimized features and weights on face-iris fusion using 46,No 3,pp 642-648 2013.
distance images», vol 137 63–75,2015. [14]Fabián Rolando Jiménez López, Camilo Ernesto Pardo
[3]Basit, A.Javed, M. Y.Anjum, M. A. «Eyelid Detection in Beainy, Oscar Eduardo Umaña Mendez « Development of a
Localized Iris Images», pp157-159, 2006. Two-Level Segmentation System for Iris Recognition Using
[4] Farmanullah Jan, Imran Usman, Shahid A. Khan, Shahzad Circular and Linear Hough Transform Algorithm» vol 5, No
A. Malik«Iris localization based on the Hough transform, a 2,pp1278-1286, 2016.
radial-gradient operator, and the gray-level intensity»124,No
23, pp5976– 5985,2013.
[15] Yarui Chen «Multimodal biometrics recognition based on
local fusion visual features and variational Bayesian extreme
learning machine», vol29, no 64, pp93-103, 2016,
[16] «Zernike Moments and Genetic Algorithm: Tutorial and
Application», vol 4, No 15, pp2217-2236, 2014.
[17] Tan, Chun WeiKumar, Ajay, «Accurate Iris Recognition
at a Distance Using Stabilized Iris Encoding and Zernike
Moments Phase Features», Vol. 23, no.9, pp. 3962-3974, 2014
[18]Yusuke Tominaga«Binary-Based Topology Optimization
of Magnetostatic Shielding by a Hybrid Evolutionary
Algorithm Combining Genetic Algorithm and Extended
Compact Genetic Algorithm», vol 49, No 5, 2093-2096, 2013.
[19] Farmanullah Jan, Imran Usman, Shahid A. Khan,
Shahzad A. Malik«Iris localization based on the Hough
transform, a radial-gradient operator, and the gray-level
intensity»vol 124, No 23, 2013, pp5976– 5985.
[20] john Daugman « New Methods in Iris Recognition » vol
5, No 37, pp 1167-1175, 2007
[21] Xing-chao ZHI « New algorithm of iris location based on
coordinate», vol 8, No 28, pp 2055-2058 , 2008.
[22]Noureddine Cherabit « Circular Hough Transform for Iris
localization», vol 2, No 5, pp114-121, 2012.
[23] Oluleye, Babatunde Leisa, Armstrong Leng, Jinsong
Dean, Diepeveen Armstrong, L Jinsong, L
Diepeveen, D «Zernike Moments and Genetic Algorithm :
Tutorial and Application», Vol. 4, no. 15, pp2217-2236, 2014.
[24] ORL database Available at:
http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.
html
[25]CASIAIrisV3Database.Availableat:
http://biometrics.idealtest.org
[26] Park, Hyun-AePark, Kang Ryoung «Iris recognition
based on score level fusion by using SVM»,Vol. 28, no. 15, pp
2019-2028, 2007
