Obtaining Stable Iris Codes Exploiting Low-Rank Tensor Space and Spatial Structure Aware Refinement for Better Iris Recognition
Kiran B. Raja, Raghavendra Ramachandra, Christoph Busch · 2019
The strength of iris recognition in terms of optimal biometric performance has been challenged by inevitable operational conditions in unconstrained scenarios. In this work we present a new approach for extracting stable iris weight maps to account for the noisy iris representation as a result of capture conditions and ineluctable segmentation errors. Traditional approaches to extract stable bits often ignore inter-code relations under the presence of multiple enrolment samples. Unlike previous works, we formulate the stable code extraction using tensor representation to exactly recover the low-rank non-noisy iris information using the multiple enrolment samples. Further, the proposed approach produces stable class specific (user specific) iris weight maps by eliminating the error bits due to sub-optimal segmentation or pupil dilation effects using spatial correspondence in a patch-wise manner. Through the set of experiments on two publicly available iris databases acquired under semi-constrained and unconstrained setting, we demonstrate the superiority for identification and verification performance over current state-of-the-art algorithms. Rank-1 identification rate on CASIAv4 distance database is achieved at 93.3% and a verification accuracy of Genuine Match Rate (GMR) of 80% at False Match Rate(FMR) of 0.0001 indicating the applicability of proposed approach in operational scenarios.