Improved Iris Recognition using Parabolic Normalization and
Multi-layer Perceptron Neural Network
A. Hilal
1,2
, B. Daya
2
and P. Beauseroy
1
1
UMR STMR - LM2S - ICD, Université de Technologie de Troyes, 12 rue Marie Curie, Troyes, France
2
IUT, Lebanese University, Saida, Lebanon
Keywords: Iris Recognition, Iris Normalization, Multi Layer Perceptron Neural Networks, Classification.
Abstract: Iris signature is considered as one of the richest, unique, and stable biometrics. This permits to an iris
identification system to identify a person even after many years from his first iris signature extraction. In
this paper we investigate a new method of iris normalization where iris features are normalized in a
parabolic function. Thus iris information close to the pupil is privileged to that close to the sclera. A
multilayer perceptron artificial neural network is then used to test the normalization effect and compare it
with classical linear normalization method. The study is tested on CASIA V3 database iris images.
Accuracy at the equal error rate operating point and receiver operating characteristics curves show better
results with the parabolic normalization method and thus propose its use for better iris recognition system
performance.
1 INTRODUCTION
Iris recognition is the process of identifying a person
using his iris signature. Among other biometric
identification techniques, it is considered as the most
recent and reliable method. This is due to the rich
features, the life time stability and the uniqueness of
the iris (Krichen 2007, Daugman 1993; 2007).
An iris recognition system can be decomposed
into 5 steps: image acquisition, iris segmentation,
normalization, coding and matching. In this paper,
we are interested in the normalization process and
more precisely on the effect of a non linear
normalization on recognition performances.
Iris images are first segmented in order to extract
and isolate the iris. Daugman’s (1993; 2007) method
has been used. Then eye lids and eyelashes are
isolated using linear Hough transform and intensity
threshold respectively. Segmented iris images are
then normalized, encoded and ready to be classified.
In what follows we give a brief review on
Daugman’s normalization technique, then we
introduce our proposed normalization method
followed by explanation on feature extraction
process, a brief matching review and after it the
matching process. Experimental procedures and
results are then reported to finish with conclusion.
2 NORMALIZATION
2.1 Daugman’s Normalization Review
Daugman (2003) approximates the iris with a
circular ring. He normalizes the iris patterns by his
‘Rubber Sheet’ called method that projects the iris
into a dimensionless rectangular shape. Intensity
pixels
,
in the Cartesian space of the
segmented iris are mapped to the Pseudo-Polar space
,
by the following equations:
,
,
,
,
,
1
,
1
(1)
(2)
(3)
where
,
and
,
are the
coordinates of the internal and external iris boundaries
respectively at angle
. r varies from 0 to 1
corresponding respectively to the internal and external
iris circular boundaries and
varies from 0 to 2.
2.2 Parabolic Normalization
According to researchers, rich iris information is
643
Hilal A., Daya B. and Beauseroy P..
Improved Iris Recognition using Parabolic Normalization and Multi-layer Perceptron Neural Network.
DOI: 10.5220/0004155406430646
In Proceedings of the 4th International Joint Conference on Computational Intelligence (NCTA-2012), pages 643-646
ISBN: 978-989-8565-33-4
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
closer to the pupil than to the sclera encouraging us
to test a method that takes into account this fact
(Krichen 2007).
While Daugman’s classical approach is to define
N equally spaced samples in each angular direction
in order to scan linearly the iris, the proposed
method experiments the normalization efficiency of
non-linearly spaced points. We redefine the spacing
of the N samples along the radius for each angular
direction of the iris. Iris pixels are normalized and
projected to the polar space according to a parabolic
function starting from the pupil boundary to the iris
external boundary. For every angle (
), samples
among varying radius (r) are picked following
always the equation (1) and following the two next
equations:
,
1
,
1
(4)
(5)
The distance between samples increases with the
distance to the pupil as illustrates figure 1.
2.3 Feature Extraction
After normalization characteristic features, the most
discriminating information of the iris, are extracted.
Many methods exist in the bibliography such as
wavelet encoding, zero-crossings of the 1D wavelet
(Boles and Boashash, 1998), Haar wavelet (Lim,
Lee, Byeon, and Kim, 2001), Laplacian of Gaussian
filters (Wildes, 1997) and finally Gabor filters
proposed by Daugman (1993) and used in our work.
Gabor filters provide a conjoint representation and
localization of iris information in space and spatial
frequency. It is constructed by modulating a sine or
cosine wave with a Gaussian. Signal decomposition
is made by implementing a quadrature pair of Gabor
filters. Real and imaginary parts are specified with a
cosine and a sine modulated respectively by a
Gaussian.A 2D Gabor filter over an image domain
,
is given by:
,
⁄
⁄
(6)
where
α,β
specify the effective width and length,
and is the filter’s angular frequency having
,
the center frequency. The filter’s phase
output represents the iris features and is used in the
matching process (Daugman 1993).
(a)
(b)
Figure 1: Illustration of parabolic normalization (a)
compared to Daugman’s linear normalization (b).
2.4 Iris Matching
2.4.1 Matching Review
Hamming distance was the first matching method
used by Daugman (1993). It s a simple Boolean that
compares images, pixel by pixel generating a match
percentage. It is not the most accurate, but its fast
computation is an essential advantage over other
metrics, such as Bayesian and Euclidean distance or
nearest feature line (Park, Lee, Smith, Park, 2003
and Yuan, Shi, 2005 and Ma, Wang and Tan, 2002).
Due to these drawbacks in classical classification
methods, use of neural networks for iris recognition
has been drawing attention (Broussard, Kennell, Ives
and Rakvic 2008 and Chen and Chu 2009). A
competitive learning vector quantization neural
network has been implemented (Lim, Lee, Byeon,
and Kim, 2001 and Cho and Kim 2006), which
learns faster than error back propagation
mechanisms. Probabilistic Neural Network and
Particle Swarm Optimization have been combined to
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
644
achieve better accuracy (Chen and Chu, 2009).
Using a rotation spreading neural network, real-time
iris recognition regardless of orientation has been
achieved by Murakami, Takano and Nakamura,
(2003). Good results are reported as well using
neural network based on VHDL prototyping by
Reaz, Sulaiman, Yasin and Leng (2004).
2.4.2 Neural Network Matching
To test our normalization method, a Multi-Layer-
Perceptron network, feed forward network with a
back propagation training method rule, is
implemented. The network has three layers: an input
layer which consists of as many neurons as there are
features in the normalized image; a hidden layer
whose number of neurons will be optimized by
checking the performance estimated with the
training set and the validation set; and an output
layer consisting of M neurons, representing each of
the M person iris signature in the database.
3 EXPERIMENTAL
PROCEDURES AND RESULTS
3.1 Iris Segmentation and
Normalization
CASIA V3-Interval database were used in this
study. We have chosen 820 images that belong to
100 person (6 to 11 images per person). The 8 bit
grayscale images are collected under near infrared
with a resolution of 320 * 280 pixels. They are
considered as good quality iris images with clear iris
texture details. Daugman’s method is used to
segment the images. After that the segmented
images are normalized according to Daugman’s
model and then according to our parabolic
normalization.
3.2 Neural Network Configuration
The input data consists of 7680 input neurons
corresponding to the number of iris features.
Unknown values related to corrupted iris templates
were replaced by a constant value of 0.01. A linear
mapping of the iris templates is performed to cover
the range of the Hyperbolic Tangent Sigmoid
function. To choose the number of neurons in the
hidden layer, network performance was tested using
a varying number from 5 to 400 neurons. Low
validation and test errors results show that 260
neurons is the best choice. Finally the output layer
consists of 100 neurons representing the 100 persons
of the database.
Weights and biases are initialized according to
the Nguyen-Widrow algorithm which distributes the
values randomly within the active region of each
neuron in the layer. To ensure convergence within a
reasonable time, experimental results reported that a
learning rate of 0.1 corresponds to the fastest
convergence conserving the same performance.
Batch training is selected as the training method
instead of online training, since the later would favor
the minimization of errors for classes having more
training patterns. As for the transfer function, it has
been found that choosing tansig for the hidden layer
and logsig for the output layer would result with the
optimal performance of the network. Cross-
validation was used to prevent over-fitting and mean
squared normalized error were found to have
superior performance than mean absolute and sum
squared error.
3.3 Parabolic Normalization
Evaluation
A total of 200 iris images (2 images per person)
were randomly selected as the train set and the rest
as the test set. The performance of the network was
used to evaluate our parabolic normalization in
comparison with Daugman’s method. The network
performance results are summarized in table 1.
Training the network takes more time and epochs
with the parabolic normalization, but compared to
Daugman’s normalization, parabolic normalization
resulted in 62.5% lower train error and 20% lower
validation error measured both on 200 images and in
30.62% lower test error measured on 620 images.
Table 1: Results of the two normalization methods.
Normalization method Parabolic Daugman
Training time 204.7 170.7
Epochs 504 411
Train error 0.0015 0.004
Validation error 0.02 0.025
Test error 0.0145 0.0209
No outer imposters are introduced in the match
process, thus only patterns from the database classes
are used. Each output node represents a distance
measurement that can be seen as a similarity score
between the iris and the corresponding class. The
maximally responding output node represents the
class membership of the input pattern.
Receiver operating characteristics (ROC) curves
ImprovedIrisRecognitionusingParabolicNormalizationandMulti-layerPerceptronNeuralNetwork
645
and accuracy at the equal error rate (EER) operating
point are used to evaluate the normalization effect.
Daugman’s normalization method resulted in
accuracy at the EER of 96.31 % while our proposed
normalization method reported a value of 97.24 %.
Figure 2 shows the ROC curves resulting for each of
the normalization methods.
Figure 2: ROC curves resulting from a parabolic
normalization compared to Daugman’s linear
normalization.
The ROC curves give best analyses of accuracy
because they present the achieved accuracy over a
range of operating points. As can be seen in figure 2,
the parabolic normalization improved accuracy at
most operating points, especially at low operating
points where significant accuracy improvements are
shown.
4 CONCLUSIONS
In this paper, we propose a novel iris normalization
method that normalizes the iris following a parabolic
function. Evaluation of the method is performed at
the matching stage using an optimized multilayer
perceptron neural network. Results compared to
Daugman’s normalization show better network
performance, more specifically, 62.5%, 20% and
30.62% lower train, validation and test error
respectively. In addition better accuracy at the EER
operating point and better ROC curves are reported
using parabolic normalization. These results show
that parabolic normalization is convenient to
represent the iris information and contribute in better
iris recognition performance.
REFERENCES
Boles, W., Boashash, B., 1998. A human identification
technique using images of the iris and wavelet
transform. IEEE Transactions on Signal Processing,
46(4).
Broussard, R., Kennell, L., Ives, R. and Rakvic, R., 2008.
An artificial neural network based matching metric for
iris identification. Proceedings of SPIE, 6812, 0-11.
CASIA-IrisV3, http://www.cbsr.ia.ac.cn/IrisDatabase.html
Chen, C., and Chu, C., 2009. High performance iris
recognition based on 1-D circular feature extraction
and PSO–PNN classifier. Expert Systems with
Applications, 36(7), 10351-10356.
Cho, S., and Kim, J., 2006. Iris recognition using LVQ
neural network. Third international conference on
Advances in Neural Networks, 3(2), 26-33.
Daugman, J., 1993. High Confidence Visual Recognition
of Persons by a Test of Statistical Independence. IEEE
Trans. in Pattern Anal. Mach. Intell, 15(11), 1148-
1161.
Daugman, J., 2007. New methods in iris recognition. IEEE
Trans. Syst., Man, Cybern. B, Cybern., 37(5), 1167-
1175.
Krichen, E., 2007. Reconnaissance des personnes par
l’iris en mode dégradé, Evry-Val Essonne University,
thesis.
Lim, S., Lee, K., Byeon, O., and Kim, T., 2001. Efficient
Iris recognition through improvement of feature vector
and classifier. ETRI Journal, 23(2), 1-2.
Ma, L., Wang, Y. and Tan, T., 2002. Iris Recognition
Using Circular Symetric Filters, 16th International
Conference on Pattern Recognition Proceedings, 2,
414- 417.
Murakami, M., Takano, H., Nakamura, K.. 2003. Real-
time Iris recognition by a rotation spreading neural
network. Annual SICE conference, 1, 283-289.
Park, C. H., Lee, J. J., Smith, M. J. T. and Park, K. H.,
2003. Iris-Based Personal Authentication Using a
Normalized Directional Energy Feature. Audio and
video based biometric person authentication
conference, 224-232.
Reaz, M. B. I., Sulaiman, M. S., Yasin, F. M., Leng, T.A.,
2004. Iris recognition using neural network based
onVHDL prototyping. International Conference on
Information and Communication Technologies: From
Theory to Applications, 463-464.
Wildes, R., 1997. Iris recognition: an emerging biometric
technology. Proceedings of the IEEE, 85(9).
Yuan, X. and Shi P., 2005. Advances in Biometric Person
Authentication. Lecture Notes in Computer Science,
3781, 135-141.
Zhu, Y., Tan, T. and Wang, Y., 2000. Biometric personal
identification based on iris patterns. 15th International
Conference on Pattern Recognition, Barcelona , Spain,
2, 801-804.
4.5 5 5.5 6 6.5 7 7.5
8
x 10
-3
0
0.005
0.01
0.015
0.02
0
.
025
False Match Rate
False Non-Match Rate
Daugman's normalization
Parabolic normalization
Daugman's normalization
Parabolic normalization
False Match Rate
x10
-
3
False Non-Match Rate
4.5
0
0
5
5.5
6
6.5
7
7.
5
8
0.005
0.01
0.015
0.02
0.025
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
646
