Using Convolutional Neural Networks for Segmentation of Iris Images

In this paper, we consider the problem of iris image segmentation. In particular, we study three convolutional neural network architectures, namely, U-Net, LinkNet, and FC-DenseNet. As a part of the study, optimal parameters for training neural networks are determined, and augmentation is used to improve the segmentation accuracy. All experiments in this paper were carried out using the open MMU Iris Database. As this dataset is not provided with the ground-truth segmentation, we performed the manual segmentation of iris images to evaluate the segmentation quality. The results obtained showed that all three neural network architectures provide high segmentation accuracy, while the best results were obtained using U-Net.