Iris-based Biometric Recognition using Modified Convolutional Neural Network

Abstract

This paper presents an iris-based biometric identification system based on a combination of Modified Convolutional Neural Network (CNN) and Softmax classifier. The system comprises: segmentation using threshold method and Hough transform, horizontal normalization of iris segmentation part and histogram equalization to form normalized images with dimensions 100×100 and 150×150. While other CNN models update the parameters in all layers, we propose a modified CNN model by splitting into two parts: feature extraction and recognition. Each normalized image is input into a CNN model, Resnet50, to extract feature vectors. Later, the extracted feature vectors are added to a Fully-Connected layers for training process and finally using Softmax layer to perform recognition task. The parameters are only updated in the fully connected layer and Softmax layer. It greatly improves computational time but still achieves high performance. The paper also performs a comparison of the identification performance of the system with different CNN models to choose the best result for iris recognition task. In addition, comparisons based on the overall accuracy performance is also conducted between the proposed system with the previous works to overview on the efficient of our approach.