Research On Iris Recognition Under Non-ideal Scenarios

Iris recognition is regarded as one of the most promising biometric recognition methods because of its security and stability. It verifies identities of human beings by comparing the iris features of captured images with those in database. After decades of research, iris recognition methods under ideal scenarios can achieve good performance. However, it remains an open problem under non-ideal circumstances where images are of pool quality and there exist serious noises and disturbances.This thesis focuses on challenging iris recognition under non-ideal scenarios. It proposes robust methods for image preprocessing, iris feature extraction and matching to deal with various disturbances such as occlusion, blur and noise. In iris preprocessing, this thesis improves the traditional method of edge detection plus Hough transformation. Specifically, the edge points detected by Canny algorithm are screened to remove noisy or erroneous points; the multiple candidate boundaries are further filtered to improve robustness and accuracy. In feature extraction, a novel iris representation method is proposed which is called ordinal measure of outer product tensor (O2PT). This method leverages high-order information of iris texture by calculating outer product tensor of raw, SIFT features. Ordinal operation is further used for improvement of robustness and matching efficiency as well as storage reduction. Moreover, this thesis makes a first attempt in modeling high-order information of periocular region by using the Fisher vector (FV) method. The two matching modalities, i.e., O2PT based iris matching and FV based periocular information, are combined for performance improvement. In matching process the mask operations are used to decrease the effects on recognition performance of eyelids occlusion and highlights.Experiments are conducted on two non-ideal databases captured under near-infrared light, i.e., CASIA-Iris-Thousand and database taken by mobiles ourselves, and non-ideal UBIris.v2 captured under visible light. The preprocessing operations including boundaries localization and eyelid detection, etc. and matching methods are evaluated and compared qualitatively and quantitatively. The results indicate the proposed methods can effectively deal with various interferences induced by non-ideal scenarios, towards accurate and robust iris recognition.