Biometric iris image segmentation and feature extraction for iris recognition

Abstract

The continued threat to security in our interconnected world today begs for urgent solution. Iris biometric like many other biometric systems provides an alternative solution to this lingering problem. Although, iris recognition have been extensively studied, it is nevertheless, not a fully solved problem which is the factor inhibiting its implementation in real world situations today. There exists three main problems facing the existing iris recognition systems: 1) lack of robustness of the algorithm to handle non-ideal iris images, 2) slow speed of the algorithm and 3) the applicability to the existing systems in real world situation. In this thesis, six novel approaches were derived and implemented to address these current limitation of existing iris recognition systems. A novel fast and accurate segmentation approach based on the combination of graph-cut optimization and active contour model is proposed to define the irregular boundaries of the iris in a hierarchical 2-level approach. In the first hierarchy, the approximate boundary of the pupil/iris is estimated using a method based on Hough’s transform for the pupil and adapted starburst algorithm for the iris. Subsequently, in the second hierarchy, the final irregular boundary of the pupil/iris is refined and segmented using graph-cut based active contour (GCBAC) model proposed in this work. The segmentation is performed in two levels, whereby the pupil is segmented first before the iris. In order to detect and eliminate noise and reflection artefacts which might introduce errors to the algorithm, a preprocessing technique based on adaptive weighted edge detection and high-pass filtering is used to detect reflections on the high intensity areas of the image while exemplar based image inpainting is used to eliminate the reflections. After the segmentation of the iris boundaries, a post-processing operation based on combination of block classification method and statistical prediction approach is used to detect any super-imposed occluding eyelashes/eyeshadows. The normalization of the iris image is achieved though the rubber sheet model. In the second stage, an approach based on construction of complex wavelet filters and rotation of the filters to the direction of the principal texture direction is used for the extraction of important iris information while a modified particle swam optimization (PSO) is used to select the most prominent iris features for iris encoding. Classification of the iriscode is performed using adaptive support vector machines (ASVM). Experimental results demonstrate that the proposed approach achieves accuracy of 98.99% and is computationally about 2 times faster than the best existing approach.