Improved iris recognition using probabilistic information from correlation filters

Abstract

The complexity and stability of the human iris pattern make it well suited for the task of biometric verification. However, any realistically deployed iris imaging system collects images from the same iris that exhibit variation in appearance. This variation originates from external factors (e.g., changes in lighting or camera angle), as well as the subject's physiological responses (e.g., pupil motion and eyelid occlusion). Following iris segmentation and normalization, the standard iris matching algorithm measures the Hamming distance between quantized Gabor features across a range of relative eye rotations. This chapter asserts that matching performance becomes more robust when iris images are aligned with a more flexible deformation model, using distortion-tolerant similarity cues. More specifically, the responses from local distortion-tolerant correlation filters are taken as evidence of local alignments. Then this observed evidence, along with the outputs of an eyelid detector, are used to infer posterior distributions on the hidden true states of deformation and eyelid occlusion. If the estimates are accurate, this information improves the robustness of the final match score. The proposed technique is compared to the standard iris matching algorithm on two datasets: one from the NIST Iris Challenge Evaluation (ICE), and one collected by the authors at Carnegie Mellon University. In experiments on these data, the proposed technique shows improved performance across a range of match score thresholds. © 2008 Springer London.