Abstract
Most prior iris recognition techniques based on the existing pipeline have already reached their limits. Therefore, this work explores the possibility of applying the deep learning technique to the field of iris recognition. We combine a novel segmentation network with a modified resnet-18 as the iris matching network. The segmentation network architecture consists of an iterative altered FCN (fully convolutional network) which contains a path of contracting layers to capture features and a symmetric upsampling path that gives precise pixel-To-pixel localization. The network not only generates visually implausible iris masks but also makes good use of data augmentation. We show that combining such networks outperforms the prior methods on several iris image datasets, including CASIA V3-interval and UBIRIS V2 datasets.