Research On Iris Recognition Algorithm Based On Deep Learning

Iris recognition technology has greater advantages than other biometric recognition technologies due to the characteristics of stability,uniqueness,anticounterfeiting,etc.Iris positioning and iris classification are two core parts of Iris recognition.Iris positioning is to find the iris boundary circle and pupil boundary circle,while iris classification is to rely on the inner and outer boundary circles obtained by iris positioning,and extract the normalized iris features to match.Therefore,an accurate and fast iris localization method and a feature extraction network with strong characterization capabilities are needed.The traditional iris positioning method relies on the edge information at the inner and outer borders of the iris,and it is easy to cause positioning failure due to interference from glasses,light spots,eyelids,etc.Considering that a person's iris and pupil are very close to a circle,and iris positioning is to find the inner and outer boundary circles of the iris,this article proposes to convert the iris positioning problem into a target detection problem required by high IoU,and output the inner and outer boundary information of the iris to improve the accuracy and speed of iris positioning.On the CASIA-Iris-Lamp and CASIA-Iris-Thousand datasets,the positioning error rates of the iris positioning algorithm based on Cascade R-CNN are 0.016 and 0.034,respectively.The OSIRIS V4.1 error rate based on the Daugman iris positioning algorithm is as high 0.4 or more.From a coarse-grained perspective,human eye parts have strong commonalities.For example,the shape and structure of the eye are consistent,the shape of the pupil is closer to a circle,and the shape of the iris.According to the characteristics of the iris image,it is not as diverse as the face image,so it is a simple but high IoU-required object detection problem.Aiming at the difficulty of obtaining annotations,this paper proposes to combine the traditional iris localization algorithm to extract annotation data.With the rapid development of deep learning in recent years,the accuracy and speed of image classification methods have been greatly improved.The essence of iris recognition is also classification,and the carrier of iris recognition is an image,so it becomes very meaningful to combine deep learning with iris recognition.Traditional iris feature extraction methods are greatly affected by eyelids,light spots,eyelashes and other factors,and their feature representation capabilities are weaker than convolutional neural networks.Therefore,in addition to the research on iris localization,this paper cuts into the iris image classification from the perspective of deep learning and proposes a multi-strategy convolutional neural network classification model for iris classification.As for the optimizer,using an optimizer composed of Radam and Lookahead can accelerate convergence on the one hand and improve accuracy on the other.According to the characteristics of the iris,the patch block input is used to strengthen the generalization of the model and improve the accuracy to a certain extent.In terms of the loss function,label smoothing is used to optimize the cross entropy loss.At the same time,center loss and soft margin triplet loss are introduced to make up for the lack of softmax loss on the distance optimization between classes.This article studies the iris positioning and iris classification.The innovations are as follows:(1)In order to solve the problem of poor positioning ability of the iris recognition method based on image processing under the interference of eyelids light spots,eyelashes,etc,it is proposed to convert the iris positioning task to high IoUrequired object detection task which outputs the inner and outer iris boundary information end-to-end;(2)In order to reduce the complexity of obtaining labeling information for object detection modeling,a combination of traditional iris positioning methods is proposed to obtain positioning information;(3)According to the different number of samples of each person in the iris data,two iris classification models mainly based on metric learning and classification,were constructed by convolutional neural networks;(4)The combined optimizer of Radam and Lookahead,multiple loss functions,and training usning patch of image were introduced to enhance representation of iris classification models.