Iris recognition based on statistical learning

This paper presents an accurate iris recognition system. The proposed iris recognition system is based mainly on statistical learning methods, and consists of four major modules. Firstly, an accurate three-step iris segmentation algorithm is proposed, during which the coarse iris centers and radius, the coarse iris boundaries, and the fine iris boundaries and centers are calculated in turn, using Least Median of Square and Linear Basis Function methods. Secondly, a two-step iris image quality evaluation process is proposed, during which the out-of-focus irises and the fail-to-segment irises are found and rejected, using our own focus function and the fine iris boundary. Thirdly, the circular area of iris is projected to a normalized rectangle with fixed height and width, using our normalization method. Finally, a many-to-many matching strategy is proposed to decrease the FAR (False Accept Rate). We tested the proposed iris recognition system on two public datasets, CASIA-Iris V3-Interval and IITD v1.0. Experimental results show that the recognition accuracy is 99.22% for CASIA-Iris V3-Interval and 99.37% for IITD v1.0, which are state-of-the-arts.