Minimal feature extraction for effective indexing of an ocular biometric system

Abstract

This thesis presents an end-to-end survey; giving the details of the various feature extraction techniques used to extract the features from the iris image, using traditional methods and deep-learning based methods. In order to extract the relevant features, suitable feature selection technique is necessary but sometimes, few redundant features may be left, which may not give the desired results. This limitation has been overcome by the use of deep-learning models. These models have the capability to extract the relevant features from the iris image by eliminating the redundant ones. In this thesis, both the approaches, including self-designed network and pre-trained network of training the deep-learning models have been applied to design an effective system. The first framework proposed a minimal feature extraction model, IrisDeepNet which can be used for effective indexing of biometric systems. An accuracy of 99.11% was achieved when the proposed model are tested on IIT Delhi iris dataset. The second framework uses pre-trained modules of GoogleNet, InceptionNet and AlexNet to design an optimum model for extracting the relevant features with the best accuracy of 99.26%. Another framework has been proposed for the design of a robust iris recognition system using noisy iris dataset, UBIRIS. The results have been compared with existing methods and it has been found that the proposed methods outperform the existing state-of-the-art methods. newline