Flexible-ICA Algorithm for a Reliable Iris Recognition

In many large scale biometric-based recognition problems, knowledge of the limiting capabilities of underlying recognition systems is constrained by a variety of factors including a choice of a source encoding technique, quality, complexity and variability of collected data. In this paper, we propose a novel iris recognition system based-on Independent Component Analysis (ICA) encoding technique, which captures both the second and higher-order statistics and projects the input data onto the basis vectors that are as statistically independent as possible. We apply Flexible-ICA algorithm in the framework of the natural gradient to extract efficient feature vectors by minimizing the mutual information of the output data. The experimental results carried on two different subsets of CASIA V3 iris database show that ICA reduces the processing time and the feature vector length. In addition, ICA has shown an encouraging performance which is comparable to the best iris recognition algorithms found in the literature.