Iris Recognition Based On Convolutional Neural Network With Few-shot Learning

Biometrics is the use of human physiological or behavioral characteristics for identity authentication and recognition.The identification method is based on the stable,unique and reliable biological characteristics of human body.Biometrics have the advantages of invariance,uniqueness,not easy to lose,security and so on.Biometrics has become an important way of security authentication,which has a very broad application prospect in security,e-commerce,finance and other fields.Human iris is randomly formed in the embryonic period,which leads to the different physiological structure of each person's iris,and the unique iris structure hardly changes in a person's life.Therefore,iris has long-term stability.In biometric technology,iris recognition has become the most ideal method of identity recognition because of its high security,anti-counterfeiting and recognition accuracy.Iris recognition technology has been studied for many years,the core of which is feature extraction.Initially,Daugman used classical Gabor filter to extract and characterize iris texture.Since then,many improved classical feature extraction algorithms have appeared,such as local binary pattern,Hilbert transform,wavelet decomposition and so on.However,the traditional hand-designed feature coder can only extract rough and single iris texture features.The deep learning technology developed in recent years can automatically extract more diversified features,and has good image perception and feature extraction ability.In this paper,convolutional neural network(CNN)is designed to extract iris image features by using deep learning method.Meanwhile,few-shot learning is used to augment iris data online to obtain better recognition performance.In this paper,an iris recognition method based on FSL and CNN is designed.By integrating online augmentation into CNN,CNN can not only learn iris texture changes caused by iris dilation,iris contraction,light change and iris rotation,but also train a deep CNN network with limited training samples,and ensure good real-time performance.This kind of targeted CNN can fully exploit the iris texture characteristics,effectively overcome the shortcomings of manual coding features in the characterization of iris texture,and automatically extract more discernible iris features.Finally,we propose an effective similarity oriented matching algorithm,which takes the first order temporal correlation coefficient as the distance metric of iris matching to improve the recognition accuracy.The proposed algorithm is evaluated on the CASIA-IrisV4-interval and CASIA-Iris-V1 databases,which are published by the Chinese Academy of Sciences.The results show that the accuracy rates of the two databases are 100%and 99.52%respectively.In addition,we also get a good result of 97.62%accuracy on the self-built SDU iris dataset.Experimental results show that the proposed method can effectively achieve iris recognition.