Iris recognition using class-specific dictionaries☆

https://doi.org/10.1016/j.compeleceng.2015.12.017Get rights and content

Abstract

In this research, the concept of class-specific dictionaries is proposed for iris recognition. Essentially, the query image is represented as a linear combination of training images from each class. The well-conditioned inverse problem is solved using least squares regression and the decision is ruled in favor of the class with the most precise estimation. An enhanced modular approach is further proposed to counter noise due to imperfect segmentation of the iris region. As such iris images are partitioned and individual decisions of all sectors are fused using an efficient fusion algorithm. The proposed algorithm is compared to the state-of-the-art Sparse Representation Classification (SRC) with Bayesian fusion for multiple sectors. The proposed approach has shown to comprehensively outperform the SRC algorithm on standard databases. Complexity analysis of the proposed algorithm shows decisive superiority of the proposed approach.

Introduction

Security is an inevitable issue of the modern world. Accordingly, for highly sensitive facilities, the traditional means of security are becoming inappropriate. Keys and cards for instance, are always prone to theft and duplication. Password protected systems could be one solution. However, due to the increasing number of utilities, users tend to have the same password for all applications, consequently making them vulnerable. In this context, biometrics are the most secure means of protection. Users are not required to memorize or carry any piece of information. Among other available biometrics such as face, speech, fingerprint and body gait, the iris has shown to be the most accurate one [1]. In particular, the texture pattern of a person’s iris is found to be unique [2] and can therefore be used as a powerful means of identification.

Daugman is considered to be the pioneer of the field [1]. He proposed, for the first time, the use of the Hamming distance classifier for the Gabor features of the iris images [3]. Since then, a variety of approaches have been employed with interesting results [4], [5], [6]. The classical concept of weighted Euclidean distance for instance, has been reported to achieve improved results [7]. In [8] the normalized correlation-based method was successfully employed, the approach is however computationally expensive owing to the direct comparison of images. The concept of the fragile bits in context of an iris code is introduced in [9]. Essentially it is shown that all bits of an iris code, derived for instance by using the Daugman’s method, are not significant. In fact the middle bands of an iris are more consistent than the inner bands. Consequently masking the bits generated by inconsistent regions of an iris (i.e. the fragile bits) has shown significant improvement over conventional methods.

A number of approaches using variants of Gabor features have been widely reported in the literature. The problem of noncooperative iris recognition is addressed using the segmentation of Gabor features in [10]. The matching of color iris images has been efficiently achieved through the hypercomplex Gabor representation [11]. Scale-invariant Gabor descriptors have been introduced in [12] to tackle the effects of hostile environments. The fusion of multiple local features has been used in combination with optimized Gabor filters in [13]. In [14] a hybrid approach has been employed. Essentially the Haar and the 1D log Gabor wavelet features have been used in combination with the SVM and HMM learning algorithms. The hybrid approach has shown to outperform the individual experts in terms of false accept rate (FAR) and false reject rate (FRR).

Recently Gabor features have been used with the state-of-the-art Sparse Representation Classification (SRC) algorithm showing an excellent performance index [15]. Essentially, a test iris image is represented as a linear combination of all training images constituting an underdetermined system of equations. The ill-conditioned inverse problem is solved using the l1-minimization. The resulting sparse vector of coefficients has ideally non-zero entries corresponding to the class of the test image.

In this research, we propose to use the intriguing concept of class-specific dictionaries rather than a single global dictionary [16]. A given probe image is simultaneously represented against all class-specific dictionaries. The resultant system of linear equations is well conditioned and the inverse problem can simply be solved using least squares regression. Finally, the decision is made in favor of the class with the minimum residual reconstruction error. With the understanding that the local analysis is more powerful compared to the holistic approach, the proposed framework is further enhanced in a modular fashion. The notion of partition-specific and class-specific dictionaries is thus introduced. As such, iris images are segmented and the individual decision for each partition is reached using the partition-specific and class-specific dictionaries. These provisional decisions are fused using the Distance-based Evidence Fusion (DEF) algorithm [16]. The DEF algorithm uses a distance metric to judge the “goodness” of a given partition thereby dynamically rejecting the non-iris partitions. It further improves the overall recognition accuracy by efficiently fusing the iris segments. To the best of our knowledge, it is for the first time that the DEF algorithm is used in the context of partition-specific and class-specific dictionaries for robust iris recognition.

The main contributions of the proposed research are: (1) least squares formulation of the iris recognition problem through the novel concept of class-specific dictionaries; (2) the use of the notion of partition-specific and class-specific dictionaries to address the problem of imperfect segmentation; and (3) the proposed DEF algorithm for the efficient fusion of the iris segments.

The rest of the paper is organized as follows: the SRC algorithm is briefly discussed in Section 2, followed by the proposed approach in Sections 3 and 4. The complexity analysis of the proposed approach is presented in Section 5. Experiments are demonstrated in Section 6 and the paper is finally concluded in Section 7.

Section snippets

Sparse representation classification for iris recognition

Let N represent the number of distinguished classes and p the number of training samples per class. A class-specific dictionary Ai is defined by, Ai=[wi(1)wi(2)…wi(p)]∈Rq×p,i=1,2,…,N such that wi(p) is a q-dimensional Gabor feature vector of the pth training iris image from the ith class. We define a global dictionary matrix A by concatenating Ai;i=1,2,…,N: A=[A1,…,AN]∈Rq×n=[w1(1),…,w1(p)|w2(1),…,w2(p)|…|wN(1),…,wN(p)]where p×N=n. Consider a test feature vector y∈Rq×1 with an unknown class

Class-specific dictionaries for reliable and rapid iris recognition

The Sparse Representation Classification (SRC) method, as discussed in Section 2, makes use of the global dictionary matrix A to develop an ill-posed linear model. Harnessing the sparsity, the inverse problem is reliably solved using the computationally expensive l1 minimization approach. In this research, we propose to use the class-specific dictionaries Ai, i=1,…,N to develop a well-conditioned system of linear equations.

Let y∈Rq×1 be a probe vector with an unknown class label, our problem is

Modular approach using class-specific and partition-specific dictionaries

View-based recognition methods, in general, follow a holistic approach. Under nominal conditions, this global processing is quite effective in developing the inter-class and intra-class models. With the practical limitations, however, the global approach may not be efficient. Degradations such as illumination, random pixel corruption, image compression [22] and contiguous occlusion, may only affect a localized portion of the image. For instance eyelashes and eyelids, in the case of iris images,

Complexity analysis

Computational efficiency is an important performance parameter for any biometric system. It is therefore imperative for a robust recognition algorithm to have a low level of complexity to ensure an inexpensive and efficient implementation. More often is the case that theoretically robust classification algorithms cannot be effectively utilized due to a high implementation cost. Owing to the high security concerns, the biometric systems in particular, need to be highly time-efficient. We

ND-IRIS-0405 database

Extensive experiments were conducted using the ND-IRIS-0405 database [25]. This database consists of 65,000 iris images from 356 subjects incorporating a number of different noise variations. The database is used as a standard testbed for performance evaluation of iris recognition algorithms. While the conventional classifiers perform well on clean iris images, they are not reliable in the presence of natural occlusions such as eyelids and eye lashes. Therefore, to thoroughly evaluate the

Conclusion

In this research a novel classification algorithm, incorporating class-specific dictionaries, was proposed for the problem of iris recognition. The proposed approach uses class-specific dictionaries to develop linear models for each subject where the resulting inverse problem is solved using linear regression and the decision is ruled in favor of the subject with the minimum reconstruction error. A modular approach was also adopted to enhance the robustness against the commonly occurring

Imran Naseem received his PhD degree from the University of Western Australia in 2010 where he is currently working as an adjunct research fellow. He is also an associate professor at Karachi Institute of Economics and Technology (KIET). He works in areas related to machine learning. Current focus of his research is on biometrics and biomedical signal processing.

References (30)

  • Wildes R.. Iris recognition: an emerging biometric technology, Proc. IEEE...
  • HollingsworthK.P. et al.

    The best bits in an iris code

    IEEE Trans Pattern Anal Mach Intell

    (2009)
  • ProncaH. et al.

    Towards non-cooperative iris recognition a classification approach using multiple signatures

    IEEE Trans Pattern Anal Mach Intell

    (2007)
  • Al-Qunaieer et al.

    Color iris recognition using hyper-complex Gabor wavelets

    Symposium on Bio-Inspired Learning and Intelligent Systems for Security

    (2009)
  • HeF. et al.

    Multiple local feature representations and their fusion based on an SVR model for iris recognition using optimized Gabor filters

    EURASIP J Adv Signal Process

    (2014)

    • Cited by (0)

    Imran Naseem received his PhD degree from the University of Western Australia in 2010 where he is currently working as an adjunct research fellow. He is also an associate professor at Karachi Institute of Economics and Technology (KIET). He works in areas related to machine learning. Current focus of his research is on biometrics and biomedical signal processing.

    Affan Aleem received his BSc(with honors) and MSc degrees from the University of Karachi and MS(Computer Science) degree from Mohammad Ali Jinnah University, Karachi. Currently, he is a PhD candidate associated with the Signal Processing Research Group at Karachi Institute of Economics and Technology (KIET). His research interests include computer vision, pattern recognition and machine learning.

    Roberto Togneri received his PhD degree in 1989 from the University of Western Australia which he joined in 1988 and is now an associate professor with the School of Electrical, Electronic and Computer Engineering. His research activities are in speech signal processing and applications in pattern recognition where he has published over 130 refereed journal and conference papers.

    Mohammed Bennamoun received his MSc degree in control theory from Queens University, Kingston, Canada, and his PhD degree in computer vision from Queens/QUT in Brisbane, Australia. He is currently a Winthrop Professor at the University of Western Australia, Australia. His research interests include machine learning control theory, robotics, object recognition, artificial neural networks, signal/image processing, and computer vision. He published more than 350 journal and conference publications.

    ☆

    Reviews processed and approved for publication by the Editor-in-Chief.

    View full text
    © 2016 Elsevier Ltd. All rights reserved.