A Visually Interpretable Iris Recognition System with Crypt Features

by Feng Shen One important barrier to the use of current iris recognition techniques in law enforcement areas such as forensics is that the iris features used by these techniques are not interpretable to human eyes. This dissertation proposes a novel iris recognition technique that determines a human’s identity based on the crypt, a visible feature on the iris. The major motivation is to aid the law enforcement applications that require human judgment in the process. The crypt is a visible feature that is relatively stable and easy for human eyes to rely on for iris recognition tasks. To verify the human-in-the-loop system design, we conducted two human-subject experiments to evaluate human perception of crypt features. The results support the applicability of using crypt features for semi-automatic iris recognition. An automatic feature detector and two crypt matchers are implemented to reduce human workload for large datasets and assist humans to make informative decisions. Our feature detector deploys a multi-scale pyramid framework in order to detect crypts of various sizes. An optimization step is then performed to improve boundary accuracies. The first matcher uses a two-stage design that approximates the overall speed of of a simple global matcher while still being able Feng Shen to locate common crypts between two irises. The second matcher uses local binary compatibility to evaluate crypt similarities, which is more robust to shape variations and partial detection errors. The performance of the crypt detector and matchers is evaluated in both identification and verification scenarios. The testing dataset is composed of 3505 images from 701 irises. It is the largest dataset that has been used to evaluate visible feature based iris recognition techniques. A final contribution of this work is the preliminary study of a semi-automatic identity verification system. Image pairs with automatically annotated crypt features and local similarity scores are presented to humans to classify each iris pair as match or non-match. The results suggest that the semi-automatic identity verification framework is feasible when the conclusion is retrieved through a voting strategy among several humans. We expect the accuracy to be considerably higher for experienced forensic examiners. To my parents, Yujun Shen and Lianqian Zhang.