Authors: Emine Krichen
Keywords: Biometric, Iris, GMM, Gabor wevelet, Daugman encoding, OSIRIS
Abstract: Several characteristics have made iris one of the most accurate biometric modality. The iris is the only internal organ of the body which is externally visible and highly stable through lifetime. It is characterized by a chaotic unique texture structure. Nevertheless, recognizing persons from their irises is not straightforward as different types of noise can be present in the image. Indeed, the iris is located behind the cornea which is a highly reflective mirror ; the resulting images will be therefore perturbed by illumination reflections.. The iris is also covered by eyelids in both its up and down part. and partly by eyelashes. This noise is very hard to detect as it has a random form and location. Blur can also be present in the images in case of non controlled acquisition conditions.
John Daugman has been a pioneer in this iris recognition field as he proposed the first iris system based on the phase coding of Gabor filter responses. Recent independent and rather large evaluation campaigns have demonstrated that the Daugman’s approach (including an efficient segmentation stage) is still nowadays the best method both in terms of performance and speed.
Nevertheless, most commercial solutions still impose constraints at the acquisition stage in order to obtain images of rather good quality.
Our aim in this thesis is to study the limits of iris recognition systems and to propose solutions when the images are acquired with few constraints, inducing a loss of quality of the images.
As a first step allowing future benchmarking, we have developed a new modular reference system called OSIRIS (Open Source for Iris) based on Daugman’s works. Our experiments show that the recognition module of OSIRIS outperforms Masek, another iris reference system publicly available (used by the NIST).
First, we were interested on studying the feasibility of iris recognition in normal light condition. In several cases, especially when irises are brown or dark, only few of the texture is apparent in the image, which generates a high ratio of failure of the reference system. We have proposed a new method based on color information in order to deal with these limitations. Our results do not allow to conclude on the possibility of using only color informaton for visible light iris recognition in degraded conditions.
In the case of Near Infrared Images, the main contribution of this thesis is the introduction of a statistical Model (Gaussian Markov Model-GMM) for characterizing good quality iris texture. This model has been used in three distinct manners :
a) for segmenting the iris in the images, when occlusions are present. Results obtained with OSIRIS show that using the probabilities of the GMM calculated on several subimages of an iris image, clearly and highly outperforms standard methods including active contour approaches.
b) for defining a novel probabilistic iris quality measure. We compare its behavior to that of other standard iris quality metrics on different types of noise which can corrupt the iris texture: occlusions, and blurring. In the case of occlusion, we compare our GMM-based quality measure to an active contour method for eyelids and eyelashes detection. In the case of blurring, comparison is made with standard methods based on gradient, Fourier Transform and wavelets. In particular, we have developed a new method able to detect blur in iris images based on wavelets. Experiments show a significant improvement of performance when our GMM-based quality measure is used instead of the classical methods above mentioned. In particular, results show that this probabilistic quality measure based on a GMM trained on good quality images is independent of the kind of ‘noise’ involved.
c) for providing a novel phase correlation-based iris matching approach able to deal with degradations in the images. Our approach is original in the sense that we do not only consider the correlation peak but also its location in different regions of the images. In a second step, the correlation-based approach is balanced with a quality metrics based on our GMM model. Experiments on different databases with different protocols show that our method improves
significantly recognition results compared to the reference systems (Masek and OSIRIS).