SHIFT AND ROTATION INVARIANT IRIS FEATURE
EXTRACTION BASED ON NON-SUBSAMPLED
CONTOURLET TRANSFORM AND GLCM
Sirvan Khalighi
1
, Parisa Tirdad
2
, Fatemeh Pak
2
and Urbano Nunes
1
1
University of Coimbra, 3030-790 Coimbra, Portugal, and Institute for Systems and Robotics (ISR-UC)
2
Department of Electrical and Computer Engineering, AZAD University of Qazvin, Qazvin, Iran
Keywords: Gray Level Co-occurrence Matrix, Iris Recognition, Non-Subsampled Contourlet Transform, SVM.
Abstract: A new feature extraction method for iris recognition in non-subsampled contourlet transform (NSCT)
domain is proposed. To extract the features a two-level NSCT, which is a shift-invariant transform, and a
rotation-invariant gray level co-occurrence matrix (GLCM) with 3 different orientations are applied on both
spatial image and NSCT frequency subbands. The extracted feature set is transformed and normalized to
reduce the effect of extreme values in the feature matrix. A set of significant features are selected by using
the minimal redundancy and maximal relevance (mRMR) algorithm. Finally the selected feature set is
classified using support vector machines (SVMs). The classification results using leave one out cross-
validation (LOOCV) on the CASIA iris database, Ver.1 and Ver.4 show that the proposed method performs
at the state-of-the art in the field of iris recognition.
1 INTRODUCTION
Iris recognition is regarded as one of the most
reliable and accurate biometric identification
technologies because of the unique, aging invariant
and non-invasive characteristics of iris. This resulted
in development of a large number of automatic iris
recognition algorithms. Daugman (Daugman, 1993)
first introduced a prototype system for automatic iris
recognition based on multi-scale Gabor wavelets and
extracted the phase information of iris textures.
Wildes (Wildes, 1997) applied a gradient-based
binary edge map and the Hough transform to detect
the iris and pupil boundaries. In (Roy et al., 2011), a
wavelet transform was applied to extract the textural
features and a genetic algorithm was employed to
select the subset of informative features.
Even though, the wavelet transform is popular,
powerful and familiar among the iris processing
techniques, it has its own limitations in capturing
directional information in images such as smooth
contours and the directional edges. This problem is
addressed by Contourlet Transform (CT) (Do and
Vetterli, 2001). In addition to multi-scale and time-
frequency localization properties of wavelets, CT
offers directionality and anisotropy. A 4-level CT
method for iris feature extraction was described in
(Li et al., 2010), in which normalized images are
partitioned into multi-scale and multi-directional
subbands. The normalized energy of subbands are
calculated as features to train a support vector
machine (SVM) classifier. Due to downsampling
and upsampling, the CT lacks shift-invariance. To
overcome this limiting factor, Cunha et al. (Cunha et
al., 2006) proposed a shift-invariant version of CT
designated non-subsampled contourlet transform
(NSCT).
In this paper a new scale, shift and rotation
invariant feature extraction method for iris
recognition in NSCT domain is proposed. After
normalizing the selected regions of interest, some
textural features are extracted from the gray level
co-occurrence matrix (GLCM) of both spatial image
and frequency subbands which resulted from NSCT
decomposition. To improve the recognition rate, the
extracted features are transformed and normalized,
then fed into the minimal redundancy and maximal
relevance (mRMR) feature selection process. Finally
the selected feature set is classified using SVMs.
2 PROPOSED APPROACH
The proposed iris recognition system includes four
470
Khalighi S., Tirdad P., Pak F. and Nunes U. (2012).
SHIFT AND ROTATION INVARIANT IRIS FEATURE EXTRACTION BASED ON NON-SUBSAMPLED CONTOURLET TRANSFORM AND GLCM.
In Proceedings of the 1st International Conference on Pattern Recognition Applications and Methods, pages 470-475
DOI: 10.5220/0003793904700475
Copyright
c
SciTePress
Figure 1: Overall System Architecture.
major phases: a) iris preprocessing, b) feature
extraction, c) feature transformation and
normalization, and d) feature selection and
classification. Figure 1 shows the architecture of the
system.
2.1 Iris Pre-processing
For the purpose of iris recognition some irrelevant
parts such as eyelid, sclera, eyelashes and pupil
should be removed. In addition, even for the iris of
the same eye, the size may vary depending on
camera-to-eye distance as well as light brightness.
Therefore, the original image needs to be pre-
processed to localize, normalize and enhance the iris
regions, and reduce the influence of the mentioned
factors.
Localization and regions of interest selection: To
locate the inner (iris/pupil) and the outer (iris/sclera)
boundaries, the following steps should be
performed:
1) Reflection removal: Specular reflections
(light spots in the eye image) can cause some
problems in the localization process. To remove the
reflections, the eye image is binarized (using a
threshold = 190). The binarized eye image is then
dilated to consider all possible affected regions.
Then the resulted mask is complemented and applied
to the eye image for marking the reflections spots.
Finally, the detected specular reflections are
inpainted” (Shah and Ross, 2009) using the 8
surrounding neighbours.
2) Pupillary boundary detection: To detect the
pupillary boundary, the eye image is first binarized
using a threshold value, M+25 (Shah and Ross,
2009) where M is the minimum fixed value of the
inpainted image. In addition to the pupil, other dark
regions of the eye image such as eyelashes fall
below this threshold value. In order to eliminate the
regions corresponded with the eyelashes, a 2-D
median filter with a 10x10 convolution mask is
applied on the binary image. This reduces the
number of candidate regions detected as a
consequence of thresholding (Shah and Ross, 2009).
The remaining regions in the median-filtered binary
image are labelled and the region with the largest
area and the smallest eccentricity is determined as
the pupil region. Finally, the pupil radius and
centroid are calculated by (1) and (2) respectively:
(4 )2pupilRadius A
π
(1)
(, )( , )
xy
C C xdA A ydA A=
∫∫
(2)
where (C
x
, C
y
) denote the center coordinates of the
pupil and A is the area of the pupil.
3) Limbic boundary detection: Before locating
the outer boundary, gamma threshold (Masek et al.,
2003) is adjusted to the iris edge map (extracted by
Canny edge detector) to enhance the iris contrast.
Then the weak edge pixels are set to zero using non-
maxima suppression; thus only the dominant edges
are extracted. Finally, the hysteresis thresholding is
applied to the image. Having the pupil center
coordinates, the radius and centre coordinates of the
iris boundary can be deduced using circular Hough
transform.
To disregard the iris regions occluded by the
eyelid and eyelashes and to avoid loss of
discriminative features, four regions of interest
(ROI) are selected:
I) right side of the iris circle, a sector between
angles –π/4 and π/4 with a radius equal to iris radius
(Figure 2 (a)). II) left side of the iris circle, a sector
between angles 4π/5 and 4π/3 with a radius equal to
iris radius (Figure 2 (a)). III) bottom side of the iris
circle, a sector between angles 4π/3 and –π/4 with a
radius of 1/2 of the iris radius (Figure 2 (b)). IV) a
disk around the pupil with a radius of 1/3 of the iris
radius to cover the pupillary area (Figure 2 (c)).
Normalization and enhancement: To compensate
several external factors such as illumination
variations and imaging distance, the partial iris
images are normalized using “Daugman Rubber
Sheet” model (Daugman, 1993).
Iris Database
Localization and
Region of interest
Normalization
and Enhancement
Test Image
Contrast, Correlation, Energy, Entropy, Dissimilarity,
Sum Average, Sum Variance, Sum of Squares Variance,
Sum Entropy, Difference Variance, Difference Entropy,
Information Measure of Correlation1&2,
Autocorrelation, Cluster Prominence, Cluster Shade,
Maximum Probability, Homogeneity, Inverse
Difference Normalized, Inverse Difference Moment
Normalized, Standard Deviation, Mean, Variance,
Energy of FFT
Feature Selection
Classification
SVM Test
SVM Training
Model
Feature Transformation and
Normalization
mRMR
Pre-processing
Feature Extraction
Result
SHIFT AND ROTATION INVARIANT IRIS FEATURE EXTRACTION BASED ON NON-SUBSAMPLED
CONTOURLET TRANSFORM AND GLCM
471
(a) (b) (c)
Figure 2: Selected regions for normalization.
(a) (b)
Figure 3: (a) Tiled normalized image. (b) Enhanced iris
image by histogram equalization and Wiener filtering.
Since the original iris image has low contrast and
may have non-uniform illumination caused by the
position of the light sources, some enhancements
need to be applied. The histogram equalization is
used to enhance the normalized iris images. The
enhancement involves tessellating the normalized
iris into 32x32 tiles (Figure 3(a)) and subjecting
each tile to histogram equalization. Then for noise-
removal, the Wiener filter is applied to each tile
(Figure 3(b)).
2.2 Feature Extraction
A reliable iris recognition system should extract
features that are invariant to scaling, shift and
rotation. The scale invariance is obtained by
unwrapping the selected iris regions into four fixed
size rectangles. To achieve shift invariance, the
enhanced images are transformed into the frequency
domain using the NSCT which is a shift-invariant
transform and can capture the geometry of the iris
texture. Finally, the GLCM is calculated on both
spatial image and NSCT frequency subbands. The
proposed method is described as follows.
Non-subsampled contourlet transform: In
contourlet transform, the Laplacian Pyramid (LP) is
first used to capture point discontinuities, and then
followed by a Directional Filter Bank (DFB) to link
point discontinuities into linear structures (Po and
Do, 2006). The overall result is an image expansion
using basic elements like contour segments, and thus
called contourlet transform, which is implemented
by a Pyramidal Directional Filter Bank (PDFB) (Do
and Vetterli, 2001). The LP decomposition at each
level generates a down sampled low pass version of
the original image, and the difference between the
original image and the prediction results in a
bandpass image. Due to downsampling and
upsampling presented in both LP and DFB,
contourlet transform is not shift-invariant. The
NSCT is built upon nonsubsampled pyramids and
nonsubsampled directional filter bank (NSDFB);
thus, it is a fully shift-invariant, multi-scale, and
multi-direction image decomposition that has a fast
implementation (Cunha et al., 2006).
Primary features: The enhanced iris image is
decomposed into 6 directions using NSDFB at 2
different scales. Afterward some textural features are
extracted from the spatial iris image and all NSCT
frequency subbands. Textural features mentioned in
Figure 1 are computed on the basis of statistical
distribution of pixels’ intensity at a given position
relative to others in a matrix of pixels called GLCM
(Haralick et al., 1973). Since the GLCM is
computed for different orientations, the rotation of
the iris can be captured by one of the matrices.
Feature extraction based on GLCM is a second-order
statistic that can be employed to analyze an image as
a texture. Although GLCM captures properties of a
texture, it cannot be directly used for further
analysis, such as the comparison of two textures;
thus numeric features which contain significant
information about the textural characteristics are
obtained from the GLCM in three different
directions (Haralick et al., 1973), (Soh et al., 1999)
and (Clausi et al., 2002).
2.3 Feature Transformation and
Normalization
The extracted features are transformed and
normalized in order to reduce the influence of
extreme values. The transformation methods applied
to each feature are described in (Becq et al., 2005).
After a thorough experimental evaluation of each
transform operator over extracted features, it was
empirically verified that the best classification
results were attained with the transform = 1/
,
where Y denotes the feature matrix, and  =


;=1,2,…,and=1,2,…, (where N and
M denote the number of subjects and features
respectively) is the transformed feature matrix.
Thereby this transform was adopted in the overall
iris recognition system. To avoid features in greater
numeric ranges dominating those in smaller numeric
ranges, each feature of the transformed matrix
Χ
is
independently normalized to the (0, 1) range by
applying
π/4
4π/5
4π/3
-π/4 -π/4
4π/3
Pu
p
illar
y
ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods
472
̅

=


−

(3)
where x
j
is a vector of each independent feature
(Aksoy et al., 2001).
2.4 Feature Selection and Classification
Larger numbers of high-dimensional feature vectors
make the classification process more complex and
less reliable due to features redundancy. To reduce
these effects the mRMR feature selector is used
(Peng et al., 2005).
Support vector machines (SVMs) (Burges, 1998)
are adopted as classifier in this study, given that
neural networks and other classifiers cannot show
reliable classification results in too noisy data.
3 EXPERIMENTAL RESULTS
The performance of the proposed algorithm was
assessed using CASIA iris image databases Ver.1
and Ver.4–Lamp (CASIA Iris Image Database).
CASIA Ver.1 contains a total of 756 grayscale iris
images, from 108 subjects, captured in two sessions
with at least one month interval. CASIA Ver.4-
Lamp was collected using a hand-held iris sensor in
one session. It contains 16213 grayscale iris images
from 411 subjects. CASIA-Ver. 4-Lamp is suitable
for studying problems of non-linear iris
normalization and robust iris feature representation
because of elastic deformation of iris texture due to
pupil expansion and contraction under different
illumination conditions.
In our experiments, a two-level NSCT
decomposition was adopted with 2 and 4 directions
for each pyramidal level respectively. Three GLCMs
were calculated on all NSCT frequency subbands
and the spatial image both in 0
o
, 90
o
and 135
o
. The
normalized iris images were decomposed by the
NSPDFB. We have used “pyrexc” and “pkva” as
NSLP and NSDFB filter in NPDFB decomposition
(Non-subsampled Contourlet Toolbox ver.1.0.0)
given their good performance. SVM-KM toolbox
(SVM and Kernel Method Toolbox) with Gaussian
kernel was used in the classification phase. The
Gaussian kernel degree and C parameters were set to
6 and 100 respectively as suggested by the best
empirical results. Experiments were carried out over
2000 images of 200 randomly selected classes, with
10 images per class and 756 images of 108 classes
for CASIA Ver.4 and Ver.1 respectively.
To estimate the accuracy of classification, leave
one out cross-validation (LOOCV) was used.
Figure 4: Comparison of selecting different iris ROI in the
localization process over the CASIAVer.4. The parameters
of method I are Ө = (0,2π), r = IrisR, method II Ө = (0,2π),
r = 1/3×IrisR and method III are Ө
Left
= (3π/4, 5π/4), Ө
Right
= (-π/4, π/4), r = IrisR.
Receiver operating characteristics (ROC) in
Figure 4 show the comparison of different iris
localization approaches on the Ver.4. Each curve is
denoted by symbols r, Ө which represent normalized
polar coordinates. Ө = (0,2π), r = IrisR refers to a
disk around the iris with iris radius, which covers the
whole iris region. Ө = (0,2π), r = 1/3×IrisR refers to
a disk around the iris with 1/3 iris radius, similar to
Figure 2 (c). Ө
Left
= (3π/4, 5π/4), Ө
Right
= (-π/4, π/4),
r = IrisR refers to a state similar to Figure 2 (a). The
results illustrate the superior performance of the
proposed approach over the other mentioned
methods in Figure 4.
From the feature extraction process, a total of
2048 features for each sample resulted. The
transformed and normalized feature matrix was fed
into the feature selection method. The number of
selected features based on the mRMR results, was
set to 357 because it provided the best mean
accuracy in a grid search. The most effective
features were correlation and homogeneity and the
least significant was maximum probability feature in
the extracted group (Figure 1). From each ROI, the
following numbers of features were selected: 68, 70,
125, 94 which correspond to Figure 2 (a) left, (a)
right, (b) and (c) respectively. These results
demonstrate that the region between (4π/3,–π/4) with
1/2 of iris radius contains more relevant features
than other regions.
Figure 5 shows the comparison of the proposed
feature extraction method using NSCT, contourlet
0 0.1 0.2 0.3 0.4 0.5 0.
6
0.9
0.92
0.94
0.96
0.98
1
False Positive Rate
Genuine Acceptance Rate
Method I
Method II
Method III
Proposed Method
SHIFT AND ROTATION INVARIANT IRIS FEATURE EXTRACTION BASED ON NON-SUBSAMPLED
CONTOURLET TRANSFORM AND GLCM
473
Figure 5: Performance of different feature extraction
methods over the CASIAVer.4.
and wavelet transforms. This diagram shows the
highest accuracy obtained by NSCT due to its
redundant structure.
Figure 6: Comparison of classification accuracy for
different frequency transformation with different numbers
of features over the CASIAVer.4.
Maximum classification accuracies of NSCT
with two different kernels, contourlet and wavelet
transforms with different number of features, are
shown in Figure 6. For the wavelet transform the
best mean accuracy of 93.40% was obtained for 322
features; however this accuracy is lower than the
case of using NSCT. For the contourlet transform
the best mean accuracy
of 96.55% was attained for
490 features, which in comparison with NSCT has a
Table 1: Performance comparison of some popular
algorithms on CASIA database Ver. 1.
Methodology
Feature Extraction
methods
Accuracy
Rate %
Daugman
(Daugman, 1993)
Gabor wavelets 100
Qi M. et al.
(Qi M. et al, 2008)
Gabor filter 99.92
Chen et al.
(Chen et al., 2009)
1-D circular profile 99.35
Poursaberi et al.
(Poursaberi et al., 2007)
wavelet
Daubechies2
99.31
Our approach without
LOOCV
NSCT and GLCM 100
Our approach by
LOOCV
(mean accuracy)
NSCT and GLCM 98.29
higher number of features. This diagram
corroborates the high performance of our approach
over the CASIA Ver.4
with the average and
maximum accuracies of 96.55% and 100%,
respectively. Furthermore, the comparison results of
the classification accurac
y over the CASIAVer.1 are
as follow. The best mean accuracy of 91.40% with
185 features and 97.35% with 565 features were
achieved for the wavelet and contourlet transforms,
respectively. The best mean accuracy of 98.29% was
achieved with the NSCT using 424 features.
Considering that there isn’t any reported result
based on CASIA Ver. 4 lamp, the proposed method
was compared with state of the art methods, just
with CASIA Ver. 1. Table 1 shows the comparison
results of the proposed method with the others.
Some accuracy results are higher than our mean
accuracy, however we highlight that our reported
results were obtained using the LOOCV method in
the testing process.
4 CONCLUSIONS
A new feature extraction method for iris recognition
based on NSCT was presented. The described
technique has some advantages over other
techniques. First, this method selects four ROIs to
make use of the most significant information in the
iris texture. Second, the extracted features are
invariant to scaling, shift and rotation, which are
important properties in the iris recognition. Third, to
reduce the effect of extreme values in the feature
matrix the extracted feature set was transformed and
normalized, which remarkably improved the
recognition rate. Fourth, mRMR was employed as a
feature selector which has proven to be one of the
most powerful and stable among known feature
0 0.02 0.04 0.06 0.08 0.1
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
False Positive Rate
Genuine Acceptance Rate
NSCT
Contourlet
wavelet
0 100 200 300 400 500 600
0
10
20
30
40
50
60
70
80
90
100
Number of features
Accuracy%
NSCT(gaussian kernel)
NSCT(Poly kernel)
contourlet(gaussian kernel)
wavelet(gaussian kernel)
150 200 250 300 350 400
97
98
99
100
ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods
474
selectors. Finally, to estimate the mean accuracy of
the proposed method LOOCV was used. The
obtained average accuracies on CASIA Ver.4 and
Ver.1 were 96.55% and 98.29% respectively, and
the accuracies without using LOOCV for both
datasets were 100% which empirically illustrate the
reliability and effectiveness of the presented method.
ACKNOWLEDGEMENTS
This work has been partially supported by the
QREN funded project SLEEPTIGHT, with FEDER
reference CENTRO-01-0202-FEDER-011530.
REFERENCES
Aksoy, S., Haralick, R. M., 2001. Feature Normalization
and Likelihood-Based Similarity Measures for Image
Retrieval. Pattern Recognition letters. 22(2001), pp.
563-582.
Becq, G., Charbonnier, S., Chapotot, F., Buguet, A.,
Bourdon, L., Baconnier, P., 2005. Comparison
Between Five Classifiers for Automatic Scoring of
Human Sleep Recordings. S.K. Halgamuge, L. Wang
(Eds.), Studies in Computational Intelligence (SCI), 4:
Classification and Clustering for Knowledge
Discovery. Springer–Verlag, pp. 113–127.
Burges, J. C., 1998. A Tutorial on Support Vector
Machines for Pattern Recognition. Data Mining and
Knowledge Discovery. Vol. 2, pp. 121-167.
CASIA Iris Image Database, http://www.cbsr.ia.ac.cn/Iris
Database.htm.
Chen, C. H., Chu, C. T., 2009. High Performance Iris
Recognition Based on 1-D Circular Feature Extraction
and PSO–PNN Classifier. Expert Systems with Appli-
cations. Vol. 36, No. 7, pp. 10351–10356.
Clausi, D. A., 2002. An Analysis of Co-Occurrence
Texture Statistics as a Function of Grey Level
Quantization, Can. J. Remote Sensing. Vol. 28, No. 1,
pp. 45-62.
Cunha, A. L., Zhou, J., Do, M. N., 2006. The
Nonsubsampled Contourlet Transform: Theory,
Design, and Applications. IEEE Trans. on Image
Processing. Vol. 15, No. 10, pp. 3089-3101.
Daugman, J. G., 1993. High Confidence Visual
Recognition of Personals by a Test of Statistical
Independence. IEEE Transactions on Pattern
Analysis and Machine Intelligence. Vol. 15, No. 11,
pp. 1148–1160.
Do, M. N., Vetterli, M., 2001. Pyramidal Directional Filter
Banks and Curvelets. IEEE Int. Conf. on Image
Processing. Vol. 3, pp. 158-161.
Haralick, R. M., Shanmugam, K., Dinstein, I., 1973.
Textural Features of Image Classification, IEEE
Transactions on Systems, Man and Cybernetics. Vol.
3, No. 6.
Li, M., Jiang, M., Han, M., Yang, M., 2010. Iris
Recognition Based on a Novel Multi Resolution
Analysis Framework. IEEE 17
th
International Confe-
rence on Image Processing, pp. 26-29.
Masek, L., Kovesi, P., 2003. MATLAB Source Code for a
Biometric Identification System Based on Iris
Patterns. The School of Computer Science and
Software Engineering, the University of Western
Australia.
Nonsubsampled Contourlet Toolbox http://www.math
works.com/matlabcentral/fileexchange/10049
Peng, H., Long, F., Ding, C., 2005. Feature Selection
Based on Mutual Information: Criteria of Max-
Dependency, Max- Relevance, and Min-Redundancy.
IEEE Transactions on Pattern Analysis and Machine
Intelligence. Vol. 27, No. 8, pp. 1226-1238.
Po, D. D. Y., Do, M. N., 2006. Directional Multiscale
Modeling of Images Using the Contourlet Transform.
IEEE Trans. on Image Processing. Vol.15, Issue 6.
Poursaberi, A., Araabi, B. N., 2007. Iris Recognition for
Partially Occluded Images: Methodology and Sensi-
tivity Analysis. EURASIP Journal on applied Signal
Processing. Vol. 2007, Issue 1.
Qi M., Lu Y., Li J., Li X., Kong T., 2008. User-Specific
Iris Authentication Based on Feature Selection.
International Conference on Computer Science and
Software Engineering.
Roy, K., Bhattacharya, P., Suen, C. Y., 2011. Towards
Non Ideal Iris Recognition Based on Level Set
Method, Genetic Algorithms and Adaptive
Asymmetrical SVMs. Engineering Applications of
Artificial Intelligence. Vol. 24, No. 3, pp. 458-475.
Shah, S., Ross, A., 2009. Iris Segmentation Using
Geodesic Active Contours. IEEE Transactions on
Information Forensics and Security. Vol. 4, No. 4, pp.
824-836.
Soh, L., Tsatsoulis, C., 1999. Texture Analysis of SAR
Sea Ice Imagery Using Gray Level Co-Occurrence
Matrices. IEEE Transactions on Geoscience and
Remote Sensing. Vol. 37, No. 2.
SVM and Kernel Method Toolbox, http://asi.insarouen.fr/
enseignants/~arakotom/toolbox/index.html
Wildes, R., 1997. Iris Recognition: An Emerging
Biometric Tech. Proc. IEEE. Vol. 85, No. 9, pp.
1348–1363.
SHIFT AND ROTATION INVARIANT IRIS FEATURE EXTRACTION BASED ON NON-SUBSAMPLED
CONTOURLET TRANSFORM AND GLCM
475