Multi-Biometric Iris Recognition System Using Consensus Between Convolutional Neural Network Architectures

Abstract

The field of biometrics represents an important domain and is currently undergoing extensive research to improve data security. Among the most used biometric modalities, we find the iris which, is specific and adequate. It is unique for each human being. Iris identification is used at international airports, sea, and land borders. In this article, we propose to use convolutional neural networks (CNN) with a consensus between the architectures (VGG, DenseNet, Resnet, Inception, MobileNet, and Xception). The goal is to classify and combine CNN architectures with the Softmax classifier focusing on consensus orientation using a voting system (Hard voting). In addition, the use of optimization techniques such as Transfer learning, Batch-normalization, Dropout, Data augmentation, Global average pooling, and Cross-validation will accelerate the learning process and avoid overfitting. Using the GPU with its parallelization will significantly speed up the processing of our program. The experiments of our approach were carried out on the MMU1 iris database with an accuracy reaching 100%.