Design of Gabor Filter Banks for Iris Recognition

Abstract

In this paper, we propose a systematic approach to design Gabor filter banks for iris feature extraction. The spectra of normalized iris images were analyzed and the most informative band (MIB) was located. When tuning the filter parameters, the pass bands of the filter banks are confined to be within the MIB. It turns out that with the MIB constraint the number of orientation parameters of the Gabor filters can be reduced to one. Furthermore, we show that the spatial locations of the Gabor filters can be arbitrarily and uniformly selected provided that the conditions of the Nyquist sampling theorem are satisfied. The main difficulty of designing filter banks for iris feature extraction is that there are too many free parameters to be determined, e.g., there will be 6N parameters to be designed for N filters. However, after analyzing the possible factors influencing the iris feature extraction process, we found that the free parameters of the Gabor filter banks can be reduced from 6N to three. By using the genetic algorithm to tune the filter parameters, the performance of our recognition system has been gradually improved. The experimental results show that the best equal error rate of our system is about 0.05% computed with the CASIA database.