Open Set Iris Recognition Algorithm Based On Deep Learning

With the development of science and technology and people’s living standards,authentication has gone deep into all aspects of life,and the security of identity information has attracted much attention.The traditional authentication methods have been difficult to meet the needs of users.Identity authentication methods based on biometrics such as face,fingerprint,palmprint and iris have attracted more and more attention.Among them,iris recognition has become the leader of many biometrics because of its uniqueness,invariance,difficult forgery and contactless acquisition.It is widely used in identity authentication.At present,most iris recognition algorithms focus on improving the accuracy of iris recognition in the case of closed set,that is,the iris to be recognized is determined as a class in the data set.However,in practical application,the iris to be recognized may not belong to existing classes in the data set,which is the recognition problem of open set data.If the iris recognition system does not have the ability to recognize the unknown class iris and judges the unknown class iris as an existing class in the system,the recognition system will be illegally invaded,resulting in the leakage of user information and serious consequences.However,the existing iris recognition algorithms do not deeply consider the iris recognition problem in the case of open sets.In view of the above problems,the deep learning is used to study the open set iris recognition algorithm in this paper.The main research contents are as follows:Firstly,in order to improve the accuracy of iris recognition system,a feature extraction network(FEN)based on multi-scale convolution feature fusion is proposed.The image features learned by convolution neural network(CNN)have excellent discrimination and generalization ability,but with the deepening of convolution layer,the size of feature map decreases.Although the number of channels is increased,some feature information,especially some edge features,will be lost.The FEN network proposed in this paper,extracts different scale feature maps generated in the convolution process,concatenates them after adaptive average pooling(Adaptive Avg Pool),realizes convolution feature fusion and fully extract iris features.Secondly,in order to make the iris recognition system distinguish unknown samples,a distance feature-based feature tuning network(FTN)is proposed.The FTN uses the iris features extracted by the FEN and the feature clustering centers of each class to generate a recognition feature containing distance parameters for iris recognition.In the verification mode,the Euclidean distance between recognition features is used for matching.In recognition mode,iris images of unknown classes can be classified by Soft Max probability and a threshold value.Finally,two loss functions including distance feature are designed for networks training.The feature clustering center needs to be used to generate recognition features by the FTN,but in the process of network training,the feature clustering center of iris will be updated frequently with the change of network parameters.In order to simplify this process,in the first stage of network training,artificially set the feature clustering center of each class and design a clustering loss.Through clustering loss training the FEN,the extracted iris features are close to the setting center,and the setting center will be used to replace the clustering center.In the second stage,to make full use of the distance information in FTN and obtain good open set recognition performance,a separation loss is designed.It makes same class iris features more concentrated and different class iris features more dispersed.since the unknown class samples do not participate in the training process of networks in this paper,the unknown class iris features extracted by the FEN will not approach the set clustering center.At the same time,the separation loss and the FTN increase the class spacing of the features extracted by the FEN,so as to strengthen the ability to identify unknown classes.In order to meet the experimental requirements of open set iris recognition,this paper constructs an open iris data set by using the existing iris data set.The experiments show that the proposed method has good open-set iris recognition performance,which not only maintains a high recognition rate for the iris of known classes,but also has good discrimination ability for the iris of unknown classes,and improves the accuracy,reliability and security of the whole iris recognition system.