Development of Robust Iris Localization and Impairment Pruning Schemes

Abstract

Iris is the sphincter having flowery pattern around pupil in the eye region. The high randomness of the pattern makes iris unique for each individual and iris is identified by the scientists to be a candidate for automated machine recognition of identity of an individual. The morphogenesis of iris is completed while baby is in mother's womb; hence the iris pattern does not change throughout the span of life of a person. It makes iris one of the most reliable biometric traits.
Localization of iris is the first step in iris biometric recognition system. The performance of matching is dependent on the accuracy of localization, because mislocalization would lead the next phases of biometric system to malfunction. The first part of the thesis investigates choke points of the existing localization approaches and proposes a method of devising an adaptive threshold of binarization for pupil detection. The thesis also contributes in modifying conventional integrodifferential operator based iris detection and proposes a modified version of it that uses canny detected edge map for iris detection.
The other part of the thesis looks into pros and cons of the conventional global and local feature matching techniques for iris. The review of related research works on matching techniques leads to the observation that local features like Scale Invariant Feature Transform(SIFT) gives satisfactory recognition accuracy for good quality images. But the performance degrades when the images are occluded or taken non-cooperatively. As SIFT matches keypoints on the basis of 128-D local descriptors, hence it sometimes falsely pairs two keypoints which are from different portions of two iris images. Subsequently the need for filtering or pruning of faulty SIFT pairs is felt. The thesis proposes two methods of filtering impairments (faulty pairs) based on the knowledge of spatial information of the keypoints. The two proposed pruning algorithms (Angular Filtering and Scale Filtering) are applied separately and applied in union to have a complete comparative analysis of the result of matching.