A Simple and Accurate CNN for Iris Recognition

Iris recognition using deep learning is a new approach in iris recognition, and many methods have been proposed so far. We consider a simple and accurate Convolutional Neural Network (CNN) as a baseline for iris recognition in contrast to the increasingly complex CNN-based iris recognition methods. In this paper, we propose a method for matching normalized iris images by dividing the iris into four regions and extracting features from each region using CNN. To reduce the influence of non-iris regions such as eyelids and eyelashes, we improve the recognition accuracy by selecting regions for training, calculating weighted matching scores based on iris regions, introducing data augmentation suitable for iris images, and introducing an attention mechanism. Through a set of experiments using the public iris image database, we demonstrate that the proposed method exhibits higher recognition accuracy than OSIRIS and other CNNs.