Algorithms for Recognition of Low Quality Iris Images

A biometric system is designed to accurately identify people based on human physiological features such as face, iris, fingerprint and voice recognition. Iris recognition is considered to have the highest identification accuracy and commercial iris recognition systems have been deployed in many applications, like passports and border control. One of the key performance limitations in iris recognition is the low image quality, including rotated iris images, partial eyelash and eyelid occlusions. We developed an algorithm to accurately detect the eyelash occlusion and eliminate these eyelash occluded regions from the recognition. The resulting detection error tradeoff curve showed an improved error rate in most of the false match (FMR) rate range. We also examined two methods of rotational invariant feature extraction: one based on the covariance features and one based on the Fourier transform magnitudes. In addition, we evaluated the errors in the localizations of circles. The system performance after correction was re-evaluated, and a marginal improvement was observed, in terms of the rank-1 identification rate, the DET curve, and the cumulative match curve. Moreover, we tested a progressive segmentation scheme. It gradually increased the area of the iris region that was segmented for the inter-class and intra-class comparisons. It showed that in a particular range of an iris image area, the system performance improves rapidly. Overall, this thesis suggests that it is possible to improve the iris recognition with low quality iris images. It also provides several research directions for future work.