Abstract: Aiming at the problem that traditional iris classification required manual design of filters to extract iris features, which was single, and usually required a large number of manual parameters adjustment, so the generalization ability was poor, we proposed an iris classification algorithm based on multiple features in residual network. On the one hand, the iris image was combined with Gabor features, on the other hand, multi-scale convolution kernels were used in the network structure, which made the learned iris features more abundant, and improved the representation ability of image features. The experimental results show that in the fixed category, using Softmax classifier for multi-classification, the classification accuracy of the algorithm in JLU iris database can be more than 98.90%, which is not lower than the network structure such as DeepIrisNet and Resnet, and the network structure of the algorithm has fewer parameters and faster learning speed.

Key words: residual network, multiple features, iris classification, lightweight