IrisCodeNet: Iris Feature Coding Network.

Using effective feature extraction algorithm to accurately represent the texture of iris is the key to iris recognition technology. Based on the particularity of the task of iris recognition, IrisCodeNet, a network model for iris feature coding, is presented in this paper. The network architecture uses an improved BasicBlock combined with a loss function AM-Softmax (additive margin softmax) that can expand the decision boundary. What's more, in order to get the best iris recognition performance, the parameters settings of AM-Softmax, the input forms of iris image preprocessed, the enhancement methods of data and the input sizes of network are studied in detail. The experimental results show that the feature extractor trained by IrisCodeNet is tested on CASIA-Iris-Thousand, CASIA-Iris-Distance and IITD iris databases. The EER (equal error rate) and TAR (true acceptance rate) evaluated are far superior to the used widely traditional algorithms. In particular, IrisCodeNet still achieves excellent recognition results without the steps of traditional iris normalization or precise iris segmentation. The visualization analysis using Grad-CAM (gradient-weighted class activation mapping) algorithm shows that the network framework pays attention to the texture of iris effectively, which proves that IrisCodeNet has a strong ability to extract features of iris. [ABSTRACT FROM AUTHOR]