Probability distribution function based iris recognition system boosted by the mean rule

Abstract

In this thesis the basic concepts of iris recognition and a new algorithm using probability distribution functions were introduced. The first section familiarized the reader with the most important facts needed to be able to grasp the idea behind different iris recognition algorithms. The second part introduced some more or less popular algorithms. Daugman‟s proposed method is the most widely used one nowadays and most of the real-life applications take advantage of it. However, it is computationally rather complex. There is also the conventional principal component analysis (PCA) that creates eigenirises out of the initial database and a method proposed by Anbarjafari et al. that uses HSI colour space and majority voting to make the decision. The third part of the thesis proposed a novel iris recognition algorithm based on the mean rule. The algorithm converts iris images from traditional RGB colour space to HSI and YCbCr and creates probability distribution functions (PDF) from channels H, S, Y, Cb and Cr for both left and right iris. Kullback-Leibler divergence is used as the metric to calculate the difference between the corresponding channels. The recognition process includes calculating KLD values for all the channels for left and right irises (i.e. there are 10 channels) and then using the mean rule to get an average of them. This means that probability of compensating errors made by some channels is quite high. In order to test the algorithm, UPOL database was used. It includes three samples for both left and right iris for 64 people. The results are described in . Even though the algorithm achieved 100% recognition rate for both left and right iris, there are theoretically several ways to enhance the performance even more like using weighted average while calculating the KLD value.