Performance Evaluation of Iris Based Recognition System Implementing PCA and ICA Encoding Techniques
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Performance Evaluation of Iris Based Recognition System
Implementing PCA and ICA Encoding Techniques
Vivekanand Dorairaj1, Natalia A. Schmid2, and Gamal Fahmy3
Lane Department of Computer Science and Electrical Engineering
West Virginia University, P.O. Box 6109, Morgantown, WV 26506
ABSTRACT
In this paper, we describe and analyze the performance of two iris-encoding techniques. The first technique is based on
Principle Component Analysis (PCA) encoding method while the second technique is a combination of Principal
Component Analysis with Independent Component Analysis (ICA) following it. Both techniques are applied globally.
PCA and ICA are two well known methods used to process a variety of data. Though PCA has been used as a
preprocessing step that reduces dimensions for obtaining ICA components for iris, it has never been analyzed in depth
as an individual encoding method. In practice PCA and ICA are known as methods that extract global and fine features,
respectively. It is shown here that when PCA and ICA methods are used to encode iris images, one of the critical steps
required to achieve a good performance is compensation for rotation effect.
We further study the effect of varying the image resolution level on the performance of the two encoding methods.
The major motivation for this study is the cases in practice where images of the same or different irises taken at
different distances have to be compared.
The performance of encoding techniques is analyzed using the CASIA dataset. The original images are non-ideal
and thus require a sequence of preprocessing steps prior to application of encoding methods. We plot a series of
Receiver Operating Characteristics (ROCs) to demonstrate various effects on the performance of the iris-based
recognition system implementing PCA and ICA encoding techniques.
Keywords: iris recognition, principal component analysis, independent component analysis, image encoding,
performance evaluation, biometrics
1. INTRODUCTION
Iris provides outstanding recognition performance when used as a biometric. Iris patterns are believed to be unique due
to the complexity of two underlying processes (1) environmental and (2) genetic that influence the generation of iris
pattern. These result in textural patterns that are unique to each eye of an individual and even distinct between twins.
Iris as a biometric has been extensively studied over the last decade. The first complete iris based recognition system
was designed and patented by J. Daugman [1]. It was followed by a number of other works. The most referred in the
literature are [2,3,4]. The systems described in these works capture and further process a single or multiple copies of
infra-red or visible light frontal view images of high resolution with no strong blur or occlusions (we call this setting
“ideal”) in order to guarantee good performance.
The literature contains a large number of publications that describe different parts of J. Daugman’s system [1-4].
The following general steps are involved:
1. Localization of region of interest – During this step, the pupil, sclera, and eyelids are segmented.
2. Normalization – Transforms a localized iris region from Cartesian coordinates to doubly dimensionless polar
coordinates.
3. Encoding – Uses 2D Gabor wavelets to encode image content that is then quantized to two levels based on phase
information of the output. The result of encoding step is presented as a binary template called “Iris Code”.
4. Matching – Matching is performed using the Hamming distance.
In practice J. Daugman’s system performs exclusively well. However, since the system is patented, there are no details
provided on how to select various parameters for its implementation. The uncertainty of specifications in the literature
1
E-mail: vivekand@csee.wvu.edu
2
E-mail: natalias@csee.wvu.edu, Telephone: (304) 293-0405 ext. 2557
3
E-mail: fahmy@csee.wvu.edu, Telephone: (304) 293-0405 ext. 3588IRIS LOCALIZATION
POLAR REPRESENTATION ENHANCEMENT AND MASKING
ICA Block PCA Block
Eigen iris 1 Eigen iris 2 Eigen iris 3
S= A ˆ X, where
−1
Eigen iris 4 Eigen iris 5 Eigen iris 6
X = PCA components
S = ICA components
ˆ −1 = Estimated unmixing
A
matrix Eigen iris 94 Eigen iris 95 Eigen iris 96
Fig. 1: Block-diagram of the system implementing PCA/ICA encoding techniques for iris.
and unavailability of “ideal” iris images for testing the system led us to reproducing step 1 and partially reproducing and
redesigning steps 2, 3, and 4 of J. Daugman’s system in our initial study of the iris [12]. Our experience has shown that
application of 2D Gabor filters produces good results. However, the performance of the entire iris recognition system is
sensitive to variations of Gabor filter parameters. The parameters of 2D Gabor filters are database dependent, meaning
that they have to be finely tuned to obtain an optimal performance each time a new database is used. The results for the
modified J. Daugman’s system constitute a baseline for performance comparison.
Most of current research works focus on redesigning the second and the third subsystems and on dealing with non-
ideal nature of iris image acquisition in practice. A large number of filter-based iris encoding algorithms [2,4,15,16]
have been designed over the past few years. These methods often demonstrate outstanding performance, especially on
datasets of good image quality. Filter parameters used for image encoding in these algorithms are typically tuned to
extract local information/features while paying less attention to the global information. In this work, we analyze the
performance of two encoding strategies (1) PCA and (2) ICA that are capable of capturing both global and local features
when applied to an iris image. The level of extracted features depends on the mode in which algorithms operate. Iris is
not a novel application for use of ICA. We are aware of a few previously published works that use ICA method for iris
image encoding [7,8,17]. However, in all these works the ICA was used in a mode of operation that extracts only local
features, as proposed by Hyvarinen [9]. The purpose of this research is not to design a perfect encoding algorithm, but
rather to explore an opportunity of using global image encoding / feature extraction algorithms to process the iris and
study the effect of combining global and local features on the performance of iris recognition system. PCA is often
used as a preprocessing step to ICA with the goal to uncorrelate data and thus simplify extraction of ICA components
[9]. To the best of our knowledge, PCA has never been analyzed as an individual iris encoding technique.
We further study the effect of varying the image resolution level on the performance of these two methods. The
major motivation for this study is the cases in practice where images of the same or different irises taken at different
distances have to be compared.
The following distinctive features characterize our approach:
1. Global iris encoding.
2. Use of Euclidean and Hamming distances to measure the performance.
3. Compensation for image rotation - takes into account the effect of head tilt.
The remaining parts of the paper are organized as follows. Sec. 2 provides a description of the two proposed
encoding approaches and explains the distinctive features of our work. Sec. 3 presents the results.
2. PROPOSED ENCODING TECHNIQUES
The general block-diagram of our system is shown in Fig.1. Below is a brief description of the two encoding blocks.2.1 Encoding Using PCA
In this section, we briefly characterize the PCA method adopted to perform iris encoding (for a details on PCA see
[6,14]). A typical PCA algorithm operates in two modes: training and testing. During the training mode, the principal
components are extracted using labeled training data. During the testing mode, the performance of the iris identification
system is evaluated.
Let M be the number of iris classes. Suppose that a training set X 1 , X 2 , K , X M , a sequence of normalized and
preprocessed iris images indexed in accordance with iris class, is available. These data are used to form the scatter
matrix
M
1
Σ=
M − 1 m =1 ∑
( X m − X )( X m − X ) T ,
where X is the empirical mean. Here we assume that images are reshaped into vector columns. Since Σ is positive
definite and symmetric, it can be decomposed using the eigenvalue decomposition method known also as the Karhunen-
Loeve expansion. Thus
Σ = QΛQ T ,
where Λ is the diagonal matrix of the eigenvalues of Σ arranged in decreasing order and Q is the orthogonal matrix
whose columns form the eigenvectors of Σ . Geometrically, the eigenvectors are the basis vectors of the transformed
~
coordinate system. In practice, the smallest eigenvalues of Σ are disregarded, and new matrix Q with vector columns
corresponding to the essential eigenvalues is formed. Thus the new transformed space has smaller dimension than the
original space (data compression concept). For iris images, as will be demonstrated in the Sec. 3, the compression is
poor.
To test the algorithm, we use additional dataset called testing dataset composed of iris images independent of the
training set. Each vector in the testing set is further projected onto the axes of the new transformed space, and the
coefficients of projections are collected in the vector of features, W . To measure the distance between two projected
iris images, we involve two distances: (i) Euclidean and (ii) Hamming. To involve the second measure, we quantize
the values of individual coefficients in the vector W to “1” or “0” if the feature value is greater than zero and less than
or equal to zero, respectively. In order to take rotation into consideration, one of two segmented and enhanced iris
images is rotated systematically on either direction (up to a few degrees) and templates are extracted for each rotated
~ ~
version of the image. Let W1 = Q T Y1 and W2 = Q T Y2 be two PCA coefficient vectors corresponding to two distinct
normalized and preprocessed iris images Y1 and Y2 from the testing set. Then to compensate for rotation the
following minimization step is performed
~ ~
min d (Q T Y1 ; Q T Y2 (θ )), (1)
θ ∈[ −θ ,θ ]
max max
~
where d (⋅) denotes the Euclidean or Hamming distance between two projected iris images and Q is the matrix
composed of eigenvectors corresponding to only essential eigenvalues. Note that for normalized images the rotation
operation reduces to the cyclic shift operation of the second image with respect to the first image.
2.2 Following up with ICA
ICA is a subspace analysis technique. It aims to find a set of independent sources that capture the underlying
randomness of the observed signals. ICA has been applied to iris analysis in [7,8,17]. In those works, randomly
selected patches of a small size from iris images form a training set. This training set is then processed using PCA to
reduce its dimensionality and decorrelate components, before applying ICA locally. In theory, whitening before ICA is
not a necessary step. Though, whitening aids estimation of independent components. The training templates available
after applying PCA form the input to the ICA block. An unmixing matrix, which represents the ICA basis vectors, is
estimated from these PCA input templates by minimizing the mutual information (a measure of dependence) between
transformed components in ICA space. Previous ICA algorithms [7,8,17] do not take rotation (image alignment) into
account. While this step seems unnecessary in [7,8,17], compensation for a rotation uncertainty is a critical step for our
non-ideal iris application. In this work, we use ICA as a follow up encoding method in anticipation that it will pick
individual fine features present in iris images and thus will improve the performance of PCA method. Unlike previous
ICA algorithms used for iris recognition, we do not divide iris images into patches during the training step but rather use
the entire iris image to estimate the unmixing matrix and further to extract ICA components.64 x 360
32 x 180
16 x 90
8 x 45
Fig. 2: Downsampling by decimation (example).
~ ~ ~
Denote by X 1 , X 2 , K , X M a sequence of preprocessed, normalized, and whitened iris images indexed by their
class. During the training mode we assume that each class is represented by a single iris image. It can be easily
generalized to a multi-image case.
ICA is a blind source-separation method. It assumes that observed data can be represented as a linear combination
of a number of independent signals. The unknowns are the mixing coefficients and the independent input signals. Let
~ ~ ~ ~
X be a matrix with vector columns given by X 1 , X 2 , K , X M and S be a matrix composed of unknown independent
input signals arranged in columns. Then ICA assumes the following forward model
~
X = AS,
where A is the unknown mixing matrix. As argued in [9], the results of linear mixing of non-Gaussian signals are
more Gaussian than the input signals. Then to estimate the mixing matrix A and one of the components of S , one has
to define a measure of non-gaussianity. One of the theoretically sound criteria is the maximization of the negentropy
given by
J = H Gaus ( A −1 X) − H ( A −1 X), (2)
where H Gaus is the entropy of the data under the assumption that data are Gaussian distributed and under the constraint
of the same covariance matrix for the distributions in H Gaus and in H (see [9] for more detailed explanation). Once the
mixing matrix and one of the input signals are estimated, the remaining input signals can be obtained by invoking the
Gram-Schmidt orthogonalization procedure. To deal with empirical case, (2) is approximated by expressions
involving empirical moments.
To test the performance of ICA method, we use testing data that are different from the training set. Similar to the
case with PCA encoding, we apply two distance measures to perform matching (i) Euclidean and (ii) Hamming
distances. This is done for the purpose of performance comparison. To overcome the effect of rotation during testing,
we project each rotated version of the two images into PCA space and then into ICA space and obtain the minimum
~ ~ ~ ~
score between the templates of the different rotated versions. Let W1 = SQ T Y1 and W2 = SQ T Y2 be two vectors of
ICA coefficients corresponding to two distinct normalized, preprocessed iris images Y1 and Y2 from the testing set.
The following minimization procedure is applied compensates for rotation
~ ~ ~
min d (SQ T Y1 ; SQ T Y2 (θ )), (3)
θ ∈[ −θ ,θ ]
max max
~
where d (⋅) denotes the Euclidean or Hamming distance between two projected iris images and Q is the matrix of
essential eigenvectors introduced in Sec. 2.2.
2.3 Effect of Varying the Resolution
In order to measure the effectiveness of the proposed algorithms based on PCA and ICA techniques, with respect to
variations in resolution, we interpolate our iris images to a lower resolution grid. Two strategies of reducing imageresolution were considered (i) down sampling and (ii) averaging. The first strategy is performed by selecting every
second pixel in both the row and column directions. The second strategy is performed by averaging the original higher
resolution image over non-overlapping blocks of size 2-by-2. Hence, in both the cases we have one point in the lower
resolution grid for every four points in the next higher resolution grid. Sec. 3 presents synthesized results. Empirical
study of the effect of varying the resolution on the performance of the iris based identification system is an ongoing
work at WVU. An example of downsampling a normalized iris image is presented in Fig. 2
3. RESULTS
All experiments were performed on the CASIA dataset provided by the Chinese Academy of Sciences [5]. The CASIA
dataset contains “non-ideal” iris images of 108 irises with 7 images per iris. The images in this dataset are strongly
occluded, blurred, and defocused. Sample images from CASIA datasets are shown in Fig. 3.
As a baseline curve, we use the receiver-operating characteristic (ROC) for modified J. Daugman’s system [12].
The results of the modified J. Daugman’s system implementation are shown in Fig. 6 (black solid line).
In our experiments, global PCA method when applied to iris images, extracted 99 essential eigenvalues. The value
of 99-th eigenvalue is approximately 15 times smaller than the value of the first eigenvalue. This resulted in a poor
compression. This also emphasizes the fact that individual iris is rich in texture.
USER 1 USER 19 USER 62
Fig. 3: Sample images from CASIA dataset.
Compensation for rotation: We first demonstrate the effect of rotation on the performance of PCA and global ICA
encoding methods. Fig. 4 and 5 display two sets of histogram distributions of genuine and imposter Euclidean distance
scores obtained using the data from the CASIA dataset. The left and the right panels in Fig. 4 show the results for iris
verification system implementing PCA encoding technique without and with compensation for rotation, respectively.
Since iris images during processing are transformed to have a pseudo-polar representation, the angle values used to
optimize the performance are measured in the number of pixels. Each image is normalized such that one pixel in the
normalized image corresponds to one degree in the original image. The range of angles used during these experiments
was set to [-10, 10]. Fig. 5 presents similar results for encoding with ICA technique. The ROC curves demonstrating
the effect of rotation are shown on the left and right panels in Fig. 6 for PCA and ICA encoding techniques,
respectively. The results are displayed both for Euclidean (ED) and Hamming (HD) distances. One can conclude that
the compensation for head tilt (rotation) leads to a substantial improvement of performance. Only for results involving
both Hamming distance and ICA, rotation operation is performed during the PCA step. The lowest scoring PCA
template is projected on ICA space.
Two scenarios: The performance of the system in Fig. 1 is further evaluated using two different scenarios. In the
first scenario, we formed two sets of iris images, training and testing, from the CASIA database. Each set consisted of
one image from 100 different irises. The training set was formed from the third image of first 100 irises, while the
testing set was formed from the second image of the same irises. In the second scenario, we used the training only to
extract eigenvectors and discarded all training set images. During testing, PCA components (for the first technique) or
ICA components (for the second technique) of each image from the test data set were obtained and compared against
PCA/ICA components of the other images in the testing set. Thus, the second scenario can be viewed as a “blind”
testing. The ROC curves for both scenarios are shown in Fig. 7. The left panel in Fig. 7 demonstrates the results when
Hamming distance is used as the matching score distance. The right panel demonstrates similar results for the casewhen
Euclidean distance is applied to calculate the matching scores. As expected, the results of testing under the “blind”
scenario are slightly degraded (both for PCA and ICA) compared to the results obtained under the first scenario.Fig. 4: The left panel shows the histogram distributions of the genuine and imposter Euclidean distances when PCA encoding
technique with no compensation for rotation is applied. The right panel presents similar results for the case when the
compensation for rotation is performed.
Fig. 5: The left and the right panels show the histogram distributions of the genuine and imposter Euclidean distances when ICA
encoding technique without and with compensation for rotation is applied, respectively.
Fig. 6: The left panel displays 4 ROC curves for PCA encoding techniques with and without compensation for rotation. The ROC
curves marked with “+” are for the case when Hamming distance is used as a matching score. The ROCs marked with “o”
describe the case when Euclidean distance is used as a matching score. The solid line marked in black corresponds to the ROC
obtained with the modified Daugman’s algorithm. The right panel shows similar results for ICA encoding.
The advantage of the “blind” scenario is in its flexibility. The coefficients forming PCA and ICA vectors need not be
stored in a user database.
Effect of varying the resolution: Fig. 8 and 9 demonstrate the effect of varying the resolution on the performance of
two proposed encoding methods employing Euclidean distance. It can be seen that the averaging strategy outperforms
the down sampling strategy. This is an anticipated result, for the averaging operation retains a larger amount of
information in the averaged image than the operation of down sampling. Note also a significant drop in performance for
some low resolutions.
Overall, our results demonstrate a potential of applying global PCA and ICA techniques for iris encoding.Fig. 7: The figure shows the ROC curves describing the performance of iris based verification system under two distinct scenarios:
(i) typical and (ii) “blind.” The results shown on the left panel are for the case when Hamming distance is used as a matching
score. The results shown on the right panel are when Euclidean distance is used.
Fig. 8: The left panel shows the effect of downsampling by averaging 2*2 neighborhood blocks, on ICA. The right panel shows the
affect of downsampling by leaving out every other pixel, on ICA.
Fig. 9: The left panel shows the effect of downsampling by averaging 2*2 neighborhood blocks, on PCA. The right panel shows the
affect of downsampling by leaving out every other pixel, on PCA.
REFERENCES
1. Daugman J. (1993) “High confidence visual recognition of persons by a test of statistical independence.” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 15(11), pp. 1148-1161.
2. Li Ma , Tan T., Wang Y. and Zhang D., “Personal Identification Based on Iris Texture Analysis”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no.12 ,pp. 1519-1533, 2003.3. Wildes R.P., “Iris Recognition: An Emerging Biometric Technology”, Proceedings of the IEEE. vol.85, No.9,
September 1997.
4. L. Masek, Recognition of Human Iris Patterns for Biometric Identification, M.S. Dissertation, The University if
Western Australia, 2003, http://www.csse.uwa.edu.au/~pk/studentprojects/libor/
5. CASIA Iris Image Database (ver. 1.0), http://www.sinobiometrics.com/casiairis.htm
6. Turk M.A., and Pentland A.P., “Face Recognition using eigenfaces” , Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 58690, Hawaii, June 1992.
7. Kwanghyuk Bae, Seung-In Noh, Jaihie Kim, "Iris Feature Extraction Using Independent Component Analysis", 4th
International Conference Audio-and Video-Based Biometric Person Authentication, Guildford, UK, June, 2003,
pp.838-844.
8. Seung-In Noh, Kwanghuk Pae, Chulhan Lee, and Jaihie Kim, "Multiresolution Independent Component Analysis for
Iris Identification", The 2002 International Technical Conference on Circuits/Systems, Computers and
Communications, 2002, Phuket, Tailand, July 2.
9. A. Hyvärinen, J. Karhunen, E. Oja, “Independent Component Analysis”, John Wiley and Sons, 2001.
10. FastICA MATLAB Package . Available online at http://www.cis.hut.fi/projects/ica/fastica
11. Bruce A. Draper, Kyungim Baek, Marian Stewart Bartlett, J. Ross Beveridge, “Recognizing Faces with PCA and
ICA”, CVIU,91(1-2):115-137, July 2003.
12. Vivekanand Dorairaj, Yash N. Shah, and Natalia A. Schmid, “Exploring and Developing Processing Techniques for
Iris Recognition System”, Biometric Consortium Conference, Crystal city, VA, USA, September 20, 2004.
13. Juha Koikkalainen, Jyrki Lötjönen: "Image Segmentation with the Combination of the PCA- and ICA-Based Modes
of Shape Variation", ISBI 2004: 149-152.
14. M. Turk and A. Pentland. “Eigenfaces for recognition”. Journal of Cognitive Neuroscience, 1(1):71–86, 1991.
15. Li Ma, Yunhong Wang, and Tieniu Tan. “Iris recognition using circular symmetric filters”. Proc. International
Conference on Pattern recognition, volume 2, pages 414–417, 2002.
16. Li Ma, Tieniu Tan, Yunhong Wang, Dexin Zhang, “Efficient Iris Recognition by Characterizing Key Local
Variations”, IEEE Trans. on Image Processing, Vol. 13, No.6, pp. 739- 750, 2004.
17. Ya-Ping Huang, Si-Wei Luo, En-Yi Chen, “An Efficient Iris Recognition System”, Proc. of the First
International Conference on Machine Learning and Cybernetics, Beijing, 4-5 Nov 2002.You can also read
