Iris Recognition Based On Deep Convolutional Features

Iris recognition refers to individual identification based on the rich texture information of the iris of the human eye.The feature extraction and modeling of iris texture is the key to the recognition algorithm.Since the powerful feature learning ability of convolutional neural networks and its successful application in multiple computer vision tasks,this paper designs two iris recognition algorithm with high recognition rates and robustness based on deep convolution features and ordinal measure modeling.Different from existing works which use fully-connected features to capture the global texture,this paper proposes to use the convolutional features for modeling local property and deformation of iris texture.Furthermore,ordinal measure modeling of convolutional features is conducted to obtain robust,computationally efficient binarized iris code(ConvOM).This paper also proposes two lightweight CNN architectures suitable for small-scale labeled iris dataset,the design also takes feature map with larger spatial size into consideration to capture more local iris texture information.Inspired by the works which embedded traditional feature modeling method in neural network for end-to-end training,this paper proposes an end-to-end recognition method based on ordinal measure embedding network(NetConvOM).Ordinal measure is embedded after the last convolutional layer of CNN model and end-to-end training is performed with triplet loss.Experimental results on three iris image benchmarks show that ConvOM surpassed traditional recognition algorithms by a large margin,especially in non-ideal scenarios with poor image quality.NetConvOM,which makes full use of the powerful learning ability of deep learning,achieved further performance improvement compared to its corresponding ConvOM,indicating the importance of joint optimization of feature learning and feature modeling.Based on above works,this paper develops a complete iris recognition application running on a smartphone with user registration and recognition functions.For quick forward inference of CNN model on mobile devices with limited computing resources,this paper studies the compression and acceleration of neural networks based on depthwise separable convolution.The parameters and computation cost of compressed model are greatly reduced,the model achieves 4x actual speedup and keeps the recognition accuracy constant.