This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2016 conference.

Iris Recognition Using SVM and ANN

Sushilkumar S. Salve1, S. P. Narote2,
Department of ENTC,
MES College of Engineering,
Pune, India.
1 2
[email protected] ,[email protected]

Abstract—In this paper we proposed an improved novel

approach to identify the person using iris recognition technique.
This approach is based on Artificial Neural Network and Support
Vector Machine (SVM) as an iris pattern classifier. Prior to
classifier, region of interest i.e. iris region is segmented using
Canny edge detector and Hough transform. Provided that the
effect of eyelid and eyelashes get reduced. Daugman’srubber
sheet model used to get normalized iris to improve computational
efficiency and proper dimensionality. Further, discriminating
feature sequence is obtained by feature extraction from
segmented iris image using 1D Log Gabor wavelet.Encoding is
done using phase quantization to get feature vectors. These Figure 1.Human Eye
binary sequence feature vectors are used to train SVM and ANN
Here, our objective is to overcome the challenge like high
as iris pattern classifier. The experimental tests are performed
over standard CASIAIrisV4 database. recognition rate of recognition system.

Index Terms— Iris, segmentation, Gabor filters, SVM,ANN II. RELETED WORK
Many researcher worked in the field of iris recognition.
I. INTRODUCTION Daugman suggest first successful working iris recognition
Due to higher need of security in public as well as in system [1]. Integro-differential operator uses first order
personal information aspects, vigorousand reliable automatic derivatives of image intensity to locate circular edges of iris
recognition of individuals has been an attractive goal [1], and upper and lower eyelids. By varying the radius r and
[2],[5],[7],[8]. In public security and information security center (x,y) position of circular contour the operator searches
domain biometrics technology plays an important role. for circular path for maximum change in pixel values. To
Biometrics is standard of comparison based on either physical achieve precise localization, the operator is applied iteratively
or behavioral traits. Fingerprint, iris pattern are physical traits with progressively reduced smoothing. In Daugman’s system,
while voice, signature includes in behavioral traits. However, by quantizing local phase angle of imaginary and real parts of
iris recognition is proven to be more beneficial as this filtered image,feature vector of 2048b were generated [1], [3].
R. P. Wilde proposed a system which uses 4 level Laplacian to
accurately identifies each individual and also distinguishes one
extract features of iris. After extracting features, normalized
from another. The human iris is thin and annular region in eye
correlation and fisher classifier is used to classify the features
located around pupil and covered by cornea, as shown in
[11].Tsai, Lin and Tao used fuzzy gray curve tracing method
figure 1, which able to provide unique and independent detail for isolate iris region. To extract local feature points and to
information of individual [3]. Random texture of protected iris evaluate dominant orientation of all detected feature points,
is mostly stable throughout the life. Iris patterns are too unique multi-resolution Gabor filters are used. Fuzzy matching
even not only differ in between identical twins but also differ algorithm is also introduce to calculate matching score [3].
in between the right and left eye. Iris recognition requires Wilde system compares feature point by using the information
comparing a register or enrolled iris template against a newly comprising the local feature and the position of each point [3].
captured iris from eye of individual [9], [16]. The iris K. Seetharaman and R. Raghupathy adopted Canny edge
recognition is applicable in identification as well as in detection method and circular Hough transform to segment the
verification mode. In identification mode recognition system iris [4]. Daugman’s rubber sheet model used to convert polar
identifies a person from entire register template by searching a coordinate to rectangular coordinate for fixed dimension iris
database for matching. Identification is ‘one to many’matching [1], [4]. After decomposing 2D normalized iris to 1D signal,
style [18], [24]. In verification mode recognition system feature will be extracted by convolving that 1D signal with 1D
authenticate a person’s claimed identity from their’s Log-Gabor filter using FFT and IFFT. These feature given to
previously enrolled patterns.Verification is ‘one to one’ phase quantization process as input, to generate binary
matching style. template i.e. iris code. Hamming distance is used for bit-wise
comparison for matching [1], [4]. Boles and Boashah were

978-1-4673-9338-6/16/$31.00 2016
c IEEE 474
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2016 conference.

analyzed iris images by zero crossing of 1D wavelet The feature extracted using 1D Log Gabor filter need to be
transform. Zero crossing representation is used to extract encoded so that binary iris code will generate and it helps in
features. Different dissimilarity functionsare used to compare classification in matching scheme. Phase quantization is used
feature vectors [10]. Ma, Tan extract local details of iris here to generate iris code.
pattern were capture by bank of spatial filter which are kernel Phase quantization assigns 11 if both real and imaginary
based on Gaussian–-Hermite moment and key local variation parts are positive, assigns 0,1 if real and imaginary parts are
method using 1D wavelet transform.Sunadopted Gaussian negative andpositive respectively, assigns 1,0 if real and
filter to calculate direction of iris image and then angle scale imaginary parts are positive and negative respectively, and
quantized into six discrete value. Yu uses multichannel Gabor assigns 00 if both real and imaginary parts are negative. On
filters to fetch key points of iris image. Related distances the basis of this logic binary stream get generated as shown in
between these key points are used to represent features [3]. matrix form,
Hyun proposed score level based iris recognition method using
SVM (Support Vector Machine) and two Gabor wavelet filters V. MULTI CLASS SVM AS IRIS PATTERN CLASSIFIERS
[25]. The code extracted from input iris image need to compare
The remainder of this paper is organized as follows. Section with previously enrolled iris code. This comparison is done
III describes the proposed approach for iris segmentation. An based on Support vector machine (SVM) classifier [21]. By
approach for feature extraction andan encoding is described in finding decision surface between two classes of input vectors
Section IV. The proposed novel matching schemes by SVM from training set. As shown in figure 2, SVM tries to
and ANN are presented in Section V and VI Sections maximize the distance of separating boundary of two classes
respectively. The experimental results are shown in Section by maximizing distance of separating planeof each of feature
VI. At last, Section VIII concludes the research work. vector.
III. IRIS SEGMENTATION
An eye image consist of iris texture and some irrelevant part
as well [13], [15], [12], [17], [23]. So, to isolate region of
interest i.e. annular iris region, the pupillary and limbic
boundaries should be detected. In this section, we describe,
proposed edge detection and inner and outer circle detection
using Canny edge detector and Hough transform [26]
respectively.
Figure 2: SVM classifier
IV. FEATURE EXTRACTION AND ENCODING If xi is input vector then g(x) is linear function given as,
To have accurate and precise recognition of person, the ‰ሺšሻ ൌ ™Ǥ š୧ ൅ „ (1)
most useful and unique information about iris pattern must be g(x) represents straight line if x is two dimensional feature
extracted. For comparison purpose only significant vector. If it is 3 dimensional then g(x) represents plane. Now,
information get encoded. Band pass decomposition [6], [17], x is greater than 3 dimension then g(x) representshyper-plane
[23], [19], [20], [22] of iris image is famous in iris recognition as shown in figure 2. Here, w is weight term which is
technique to create biometric template. In this paper, perpendicular to hyper-plane. Orientation of hyper-plane is
normalized iris pattern is convolved with 1D Log–Gabor given by w. Also, b is bias term which indicates position of
wavelet for actual implementation of feature extraction. hyper-plane. Hyper-plane divides D dimensional space into
Number of 1D signal are recovered by breaking 2D two half spaces. In two class problem classification rule if
normalized iris patterns. Then using FFT and IFFFT, ‰ሺšሻ ൌ ™Ǥ š୧ ൅ „ ൐ Ͳ then, xi belong to class c1. If ‰ሺšሻ ൌ
convolution is done in between 1D signal and 1D Log-Gabor ™Ǥ š୧ ൅ „ ൏ Ͳ then, xi belongs to class c2. While classifying
wavelet. Gabor wavelets are able to provide the best any input vector we assign class belongingness yi=േͳ.If
representation of signal in both space and spatial frequency. ‫ݔ‬௜ ߳ܿଵ then ‫ݕ‬௜ = േͳ and if ‫ݔ‬௜ ߳ܿଶ then ‫ݕ‬௜ =-1. Hence‫ݕ‬௜ ሺ‫ݓ‬Ǥ ‫ݔ‬௜ ൅
Sine/cosine wave get modulated with Gaussian to construct ܾሻ ൐ Ͳ is always positive whether ‫ݔ‬௜ belongs to ܿଵ orܿଶ . Now
Gabor filter. Generally, a sine wave is much perfectly for stability of each feature vector we consider a margin
localizes in frequency than space. After sine modulated with ሺߛሻ.Therefore,‫ݓ‬Ǥ ‫ݔ‬௜ ൅ ܾ ൐ ߛ. This margin is nothing but a
Gaussian provide space localization but at same time it loses measure of distance of ‫ݔ‬௜ from hyper-plane. So, distance of
localization in frequency. The signal decomposition carried by ௪Ǥ௫ା௕
hyper-plane from that point ‫ݔ‬௜ is given by ԡ௪ԡ ൒ ߛ. By
Gabor filters. The drawback of Gabor filter is that, whenever
the bandwidth exceeds one octave even symmetrical filter will proper scaling we can setߛǤ ԡ‫ݓ‬ԡ ൌ ͳ. Therefore,
have DC component. We can get the zero DC component w.x+b=1.Now classification rule becomes,
using Gabor filter for any bandwidth which is nothing but ™Ǥ š୧ ൅ „ ൒ ͳ ՜ ‫ݔ‬௜ ߳ܿଵ ܽ݊݀™Ǥ š୧ ൅ „ ൑ ͳ ՜ ‫ݔ‬௜ ߳ܿଶ (2)
Gaussian over logarithmic scale, which is known as Log– If š୧ is support vector then ›୧ (w.š୧ +b=1).To design SVM we
௪Ǥ௫ା௕
Gabor wavelet [11], [23], [24], [25], [26] filter. need margin as large as possible. So, margin will be ԡ௪ԡ ൒ ߛ.
To get largest possible margin two point take into

475
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2016 conference.

consideration. (1) Minimize ԡ‫ݓ‬ԡ (2) maximize b. This

problem can be convert to un-constraint optimization problem
by using Lagrangian multiplier.
ଵ
‫ܮ‬ሺ‫ݓ‬ǡ ܾǡ ߙሻ ൌ ԡ‫ݓ‬ԡଶ െ σே ௜ୀଵ ߙ௜ ሺ‫ݕ‬௜ ሺ‫ݔ‬௜ Ǥ ‫ ݓ‬൅ ܾሻ െ ͳሻ(3)
ଶ
Where, ߙ௜ is Lagrangianmultiplier. To optimize that take
partial derivatives of above Lagrangian with respect to w and
b. So we get,
డ௅ డ௅
ൌ σ௠ ௠
௜ୀଵ ߙ௜ ‫ݕ‬௜ ൅ ͲAndడ௕ ൌ ‫ ݓ‬െ σ௜ୀଵ ߙ௜ ‫ݕ‬௜ ൌ Ͳ
డ௕
Finally, we get
ଵ Figure.3: A feed forward neural network
‫ ܮ‬ൌ σ௠ ௜ୀଵ ߙ௜ െ ଶ σ ߙ௜ ߙ௝ ‫ݕ‬௜ ‫ݕ‬௝ ሺ‫ݔ‬௜ Ǥ ‫ݔ‬௝ ሻ (4)
To, become SV ߙ௜ must be very large. It means ߙ௜ shows the Feed forward neural network using Levenberg Marquardt rule:
effect of feature vector on decision boundary. In order to In our work, we also apply the feed forward neural network
obtain a non-linear decision boundary, we replace the inner using Levenberg-Marquardt rule (FFLM) proposed for iris
product ൫‫ݔ‬௜ Ǥ ‫ݔ‬௝ ൯ with a nonlinear kernel ‫ܭ‬൫‫ݔ‬௜ Ǥ ‫ݔ‬௝ ൯ܽ݊݀݃݁‫ݐ‬ pattern classification
Let us compare the feed forward network figure 4 .The
ˆሺšሻ ൌ σ୒
୧ୀଵ ›୧ Ƚ୧ ሺšǤ š ୧ ሻ ൅ „ (5) network input to uniti in the layer k+1 is defines as,
ௌ௞
The simple idea behind the nonlinear SVM [9] is to use a ݊௞ାଵ ሺ݅ሻ ൌ ෍ ‫ ݓ‬௞ାଵ ሺ݅ǡ ݆ሻܽ௞ ሺ݆ሻ ൅ ܾ ௞ାଵ ሺ݅ሻ
kernel function ‫ܭ‬ሺ‫ݔ‬Ǥ ‫ݔ‬௜ ሻ to map the data x from the input ௝ୀଵ
space to the new higher dimensional feature space on which And the output of the unit i is calculated as
the mapped data points become linearly separable. The hidden ܽ௞ାଵ ሺ݅ሻ ൌ ݂ ௞ାଵ ൫݊௞ାଵ ሺ݅ሻ൯
units on the hidden layer are characterized by the kernel The system equations for M layer network can be presented in
function adopted by SVM. The two basic kernels used in this matrix form as follows
work are, ܽ଴ ൌ ‫݌‬
Radial Basis Function (RBF) kernel: ܽ௞ାଵ ൌ ݂ ௞ାଵ ሺܹ ௞ାଵ ܽ௞ ൅ ܾ ௞ାଵ ሻǡ ݇ ൌ Ͳǡͳǡ ǥ ǥ ǤǤǤ Ǥ ǡ ‫ ܯ‬െ ͳ
ԡ௫ି௫೔ ԡమ
‫ܭ‬ሺ‫ݔ‬ǡ ‫ݔ‬௜ ሻ ൌ ݁‫ ݌ݔ‬ቄെ ቅ (6) The neural network learns the associations between a specified
ଶఙ మ
௫Ǥ௫೔ ଶ set of input-output pairs represented as,
Polynomial kernel:‫ܭ‬ሺ‫ݔ‬ǡ ‫ݔ‬௜ ሻ ൌ ቀͳ ൅ ቁ (7)
ఙమ

VI. FEED FORWARD NEURAL NETWORK AS IRIS PATTERN

CLASSIFIER
Neural networks are biologically inspired and statistically
based technique. The networks are made up of processing
units called neurons. The structure of neuron is the basic
Figure 4: A feed forward network
building block of the network. The neuron calculates the
weighted sum over the output of neuron linked to its input. The performance index of the neural network is measured as
The activation function is applied to the weighed sum to ଵ ் ் ଵ
determine the neuron’s output. The sigmoid function is used as ܸ ൌ σொ௤ୀଵ൫‫ݐ‬௤ െ ܽ௤ ൯ ൫‫ݐ‬௤ െ ܽ௤ ൯ ൌ σொ௤ୀଵ ݁௤் ݁௤ (9)
ଶ ଶ
an activation function and is defined as.
݃ሺ‫ݕ‬ሻ ൌ ݂ሺ‫ݕ ் ݓ‬ሻ ൌ
ଵ Here, ‫ ݍ‬௧௛ input of thr network is ‫ି݌‬௤ and ܽି௤ ெ
denotes the
೅ 
ଵାୣ୶୮ሺି௪ ௬ሻ
(8)
output of the network. Then the error for the ‫ ݍ‬௧௛ input is
ெ
Here, y is the vector z improved with constant value 1, estimated as ݁௤ ൌ ‫ିݐ‬௤ -ܽି௤ .The gradient (steepest) descent rule
where the vector z signifies the set of specific sample of an is used for standard backpropogation algorithm. The LM
object called measurement vector. The w is the weight vector, algorithm is an approximation of the Newton’s method. Let us
and the specific weight which links to the constant value 1 in z consider the function ܸሺ‫ݔ‬ሻ which is required to be minimized
is denoted as the bias weight. In a neural network, several with respect to the parameter vector x, and then the Newton’s
layers of different number of neurons are connected as shown method is defined as,
in figure 3. ο‫ ݔ‬ൌ െሾ‫׏‬ଶ ሺ‫ݔ‬ሻሿିଵ ‫ܸ׏‬ሺ‫ݔ‬ሻ (10)
Where, ‫׏‬ଶ ሺ‫ݔ‬ሻ denotes Hessian matrix and ‫ܸ׏‬ሺ‫ݔ‬ሻ is the
gradient. Suppose that ‫ܸ׏‬ሺ‫ݔ‬ሻ is a sum of squares function
ܸሺ‫ݔ‬ሻ ൌ σே ଶ
௜ୀଵ ݁௜ ሺ‫ݔ‬ሻ (11)
Then we can show that,
‫ܸ׏‬ሺ‫ݔ‬ሻ ൌ ‫ ்ܬ‬ሺ‫ݔ‬ሻ݁ሺ‫ݔ‬ሻ

476
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2016 conference.

‫׏‬ଶ ܸሺ‫ݔ‬ሻ ൌ ‫ ்ܬ‬ሺ‫ݔ‬ሻ‫ܬ‬ሺ‫ݔ‬ሻ ൅ ܵሺ‫ݔ‬ሻ transform to detect eyelids. Thresholding is used in order to
Where, J(x) presents the Jacobian matrix calculated as follows, suppress eyelashes. Here, onwards we show segmentation
డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ results on each dataset. While implementing the segmentation
‫ ۍ‬డ௫భ డ௫మ డ௫೙ ‫ې‬ technique we take 1000 images of above five datasets
‫ێ‬డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ ‫ۑ‬
‫ܬ‬ሺ‫ݔ‬ሻ ൌ ‫ ێ‬డ௫ ǥ‫ۑ‬and containing 200 images of each. So the properly or correctly
భ డ௫మ డ௫೙
‫ێ‬డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ డ௘ಿ ሺ௫ሻ ‫ۑ‬
segmented iris images are as shown in Figure 5.
‫ ۏ‬డ௫భ డ௫మ డ௫೙ ‫ے‬
ே

ܵሺܺሻ ൌ ෍ ݁௜ ሺ‫ݔ‬ሻ‫׏‬ଶ ݁௜ ሺ‫ݔ‬ሻ

௜ୀଵ
We assume that ܵሺ‫ݔ‬ሻ ൎ Ͳǡ for the Gauss- Newton method and
then the
ο‫ ݔ‬ൌ ሾ‫ ்ܬ‬ሺ‫ݔ‬ሻ‫ܬ‬ሺ‫ݔ‬ሻሿିଵ ‫ ் ܬ‬ሺ‫ݔ‬ሻ݁ሺ‫ݔ‬ሻ
Now the Gauss- Newton method is modified as follows, Figure 5: Segmentation result on CASIA V4 Iris Interval
ο‫ ݔ‬ൌ ሾ‫ ்ܬ‬ሺ‫ݔ‬ሻ‫ܬ‬ሺ‫ݔ‬ሻ ൅ ߤ‫ܫ‬ሿିଵ ‫ ்ܬ‬ሺ‫ݔ‬ሻ݁ሺ‫ݔ‬ሻ (12) The overall segmentation results are summarized as follow,
Which is treated as LM modification,\. Here, the parameter ߤ TABLE II
plays an important role. The parameter ߤ is divided by ߚ if the COMPARISON OF VARIOUS METHODS USED FOR ESTIMATION
step reduces V(x) and on the other hand, if step increases Data Set Total no. of Segmented Segmentation
V(x), then ߤ is multiplied byߚ. If the value ߤ is large tehn the Images Images Accuracy(%)
algorithm becomes gradient decent and while for small ߤ the CASIA V4- 200 200 100
algorithm become Gauss-Newton. Thus, we can summarize Iris Interval
the procedure of LM algorithm as follows, CASIA V4- 200 194 97
1. The output corresponding to all input to the network and Iris Lamp
the errors ݁௤ ൌ ‫ݐ‬௤ -ܽି௤ெ
are computed and the sum of CASIA V4- 200 197 98.5
squares of errors over all inputs V(x) is calculated. Iris Syn
2. The Jacobian matrix J(x) is estimated CASIA V4- 200 185 92.5
3. Solve above equation with ο‫ ݔ‬inorder to calculate ο‫ݔ‬ Iris Thousand
4. The sum of squares of errors is re-estimated using x+ο‫ݔ‬ CASIA V4- 200 183 91.5
.If the newly computed sum of squares is smaller than that Iris Twins
calculated in step 1, then ߤ is decreased by ߚ and go back Total 1000 959 95.9
to step 1. If the sum of squares is greater than that in step
1, ߤ is increased by ߚ and go to step 3. II Classification Results
5. It is assumed that the algorithm is converged if the norm In this research work we use Support vector machine and
of gradient is less than some predetermined value, or ANN. In training phase we pass all 1000 images and take
when the sum of squares is decreased to some error goal. respective segmentation result and features are stored in the
form of template.
VII. EXPERIMENTAL RESULTS TABLE III
CLASSIFIXATION RESULTS USING ANN
In order to evaluate performance of SVM kernels and ANN
we use freely available CASIA database which most widely Total No. of Correctly Accuracy
Images (A) Classified B/A*100
used for iris biometric purposes. It was not possible to use all Image(B)
of the eye images from each database, since perfect
CASIA V4- Iris
segmentation success rates were not attained. Instead a sub-set
Interval 200 190 95
of each database was selected, which contained only those CASIA V4- Iris
images that were segmented successfully. Lamp 200 188 94
TABLE I
DETAILS OF EACH SUB-SET EYE IMAGES USED
CASIA V4- Iris
Syn 200 184 93
Name of Data Set Total no. of images used CASIA V4- Iris
CASIA V4- Iris Interval 200 Thousand 200 180 90
CASIA V4- Iris Lamp 200
CASIA V4- Iris Syn 200 CASIA V4- Iris 200 181 90.5
CASIA V4- Iris Thousand 200 Twins
CASIA V4- Iris Twins 200 Total 1000 959 92.5
Total 1000 Radial basis function kernel and polynomial kernel are two
kernels of SVM which we are implemented. The classification
I Segmentation Results results given by using RBF kernel and polynomial kernel
While implementing segmentation technique, we use circular compared with ANN is as shown follow,
Hough transform to localize iris boundary and linear Hough

477
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2016 conference.

TABLE IV [5] K.Seetharaman and R. Raghupathy. “Iris recognition for personal

CLASSIFIXATION RESULTS USING SVM KERNELS identification system”. Procedia engineering (38) page 1531-1546,
2012.
Correctly Classified Accuracy
Images (B) B/A *100 [6] J. G. Daugman . “Biometric personal identification system based on iris
Total analysis,” U. S. Patent 5,291,560,1994.
No. of [7] Wildes et al.. “Automated ,noninvasive iris recognition system and
Polynomial RBF Polynomial RBF method”.U.S. patent 5,572,596, 1996
Images
Kernel Kernel Kernel Kernel
(A) [8] R.P. Wildes . “Iris Recognition: An Emerging Biometric Technology”,
Proceedings of the IEEE, vol.85, pp.1348-1363,1997.
CASIA [9] Kaushik Roy, Prabir Bhattacharya, Chin Y.Suen . “Towards nonideal
V4- Iris 200 198 200 99 100 iris recognition based on level set method , genetic algorithms and
Interval adaptive asymmetrical SVMs”.Engineering application of artificial
CASIA intelligence (24) page no.458-574, 2011.
V4- Iris 200 192 194 96 97 [10] N. Ritter . “Location of the Pupil-Iris Border in Slit Lamp Images of the
Lamp Cornea”, Proceedings of the International Conference on Image
Analysis and Processing, 1999
CASIA
V4- Iris 200 193 197 96.5 98.5 [11] S. Lim, K. Lee, O. Byeon, and T.Kim. “Efficient Iris Recognition
through Improvement of Feature Vector and Classifier”, ETRI Journal,
Syn vol. 23, no.2, pp. 61-70, 2001.
CASIA
[12] Zhaofeng He, Zhenan Sun. “Towards Accurateand Fast Iris
V4- Iris 200 181 185 90.5 92.5 Segmentation for Iris Biometrics ”IEEE Transactions on Pattern
Thousand Analysis and Machine Intelligence, Vol.31, No.9, Sept 2009
CASIA [13] J. Cui, Y. Wang, T. Tan, L. Ma, and Z. Sun . “A Fast and Robust Iris
V4- Iris 200 182 183 91 91.5 Localization Method Based on Texture Segmentation”, SPIE Defense
Twins and Security Symposium, vol. 5404, pp. 401-408, 2004.
Total 1000 946 959 94.6 95.9 [14] W. Kong and D. Zhang . “Accurate iris segmentation based on novel
reflection and eyelash detection model”, Proceedings of 2001
VIII. CONCLUSIONS International Symposium on Intelligent Multimedia, Video and Speech,
2001.
The segmentation results are mainly depends on how [15] D. M. Monro, S. Rakshit, and D. Zhang , “DCT Based Iris Recognition”,
properly preprocessing of eye images are carried out. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
Preprocessing may affected by lightning environment, image 29, no. 4, 2007.
capture distance and other noise factors. Once preprocessing [16] D. Field. “Relations between the statistics of natural images and the
execute well then segmentation result will be much better. response properties of cortical cells”, Journal of the Optical Society of
America.1987.
Overall classification results are depends on segmentation
[17] Ma L., Tan T., Wang Y., and Zhang D.. Personal identification based on
accuracy. As proper segmentation achieved then surely iris texture analysis. IEEE Trans. Pattern Anal. Mach. Intelligence,
classification results will be proper. Hough transform play an 25(12):1519 – 1533, 2007
important role for locating the region of interest i.e. iris. Gabor [18] MakramNabti, LahouariGhouti and Ahmed Bouridane,” An effective
filter provide the feature that will encode by phase and fast iris recognition system based on a combined multiscale feature
quantization method. Feature extraction and encoding is very extraction technique”, Elsevier-Pattern Recognition 41, 868-879, 2008
crucial as classification is totally based on encoded features. [19] C. Sanchez-Avila, R. Sanchez-Reillo, and D. De Martin- Roche “Iris-
Based Biometric Recognition Using Dyadic Wavelet Transform”, IEEE
The main contribution of this research work is hide in AESS System Magazines, vol. 17, no. 10, pp. 3-6, 2002.
accuracy of classification results. To evaluate the classification [20] Y. Zhu, T. Tan, and Y. Wang “Biometric Personal Identification Based
result we implement RBF kernel and polynomial kernel along on Iris Patterns”, Proceedings of the 15th International Conference on
with ANN. The classification accuracy given by RBF kernel is Pattern Recognition, vol. 2, pp. 2801-2804, 2000.
95.9%, polynomial kernel is 94.6% and by ANN92.5%. Based [21] Kaushik Roy, Prabir Bhattacharya and Ramesh Chandra Debnath.
“Multi-Class SVM Based Iris Recognition”, international conference on
on these results we can conclude that RBF kernel is slightly computer and information technology, 2007
more accurate in terms of classification accuracy. [22] L. Ma, Y. Wang, and T. Tan ,“Iris Recognition Based on Multichannel
Gabor Filtering,” Proc. Fifth Asian Conf. Computer Vision, vol. I, pp.
REFERENCES 279-283, 2002,
[1] J. Daugman “How iris recognition works”, IEEE Trans. CSVT, vol. 14, [23] A. Poursaberi and B.N. Araabi. “Iris Recognition for Partially Occluded
no. 1, pp. 21 – 30. 2004. Images: Methodology and Sensitivity Analysis”, EURASIP Journal on
[2] J. Daugman “The importance of being random: statistical principles of Advances in Signal Processing, vol 2007.
iris recognition”, Journal of pattern recognition society, page 279- [24] A. Poursaberi and B.N. Araabi “A Novel Iris Recognition System Using
291,2003.. Morphological Edge Detector and Wavelet Phase Features”, GVIP (05),
[3] Chung-Chih Tsai, Heng-Yi Lin, Jinshiuh Taur ad Chin-Wang Tao “Iris No. V6, pp. 9-15, 2005.
recognition using possibilistic fuzzy matching on local feature”. IEEE [25] W. Boles and B. Boashash “A human identification technique using
Transaction on systems, man and cybernatics.Vol.2.No.1, 2012. images of the iris and wavelet transform”, IEEE Transactions on Signal
[4] J.Daugman,.“New methods in iris recognition”. IEEE transaction on Processing, vol. 46, no. 4, 1998..
systems, man and cybernatics.Vol.37.No.5. 2007. [26] L. Ma, T. Tan, Y. Wang, and D. Zhang “Efficient Iris Recognition by
Characterizing Key Local Variations”, IEEE Trans. Image Processing,
vol 13, no.6, pp. 739-750, 2004.

478