The Performance of Deep Neural Networks in Deformed Iris Recognition System

Abstract

Iris recognition system is a powerful tool for person identification based on their special iris traits which are unique for each individual. Besides hand-crafted techniques, modern deep learning algorithms can be implemented as a good feature extractor for iris recognition system. Deformed iris texture due to different pupil dilation of the same eye can cause negative impact on iris recognition performance. Hence in this paper, we study the effectiveness of deep learning algorithms on extracting deformed iris features from the normalized iris images. We compared the performance of 17 available deep learning algorithms which followed by a multi-class Support Vector Machine (SVM) algorithm to perform classification. We then utilized different epoch number on the model until a good accuracy is achieved. We also extracted features using different type of layers in order to identify which type of layers could extract good features. Simulation results of 92% by Darknet-19 on CASIA-Iris-Lamp dataset reveal the effectiveness of deep learning algorithms on extracting irregular features of deformed iris.