Research On Iris Recognition Algorithm Based On Deep Learning

With the development of the times,biometric-based identity recognition gradually replaces traditional identity authentication methods and is widely used in people's daily lives.Among the many biological features,iris recognition has the highest stability and security,so it has a wide range of application scenarios and high research value.In view of the shortcomings of the existing iris recognition algorithms,deep learning is used to improve iris segmentation and recognition.Feature fusion network based on attention mechanism is used for iris segmentation and residual twin network is used for iris recognition.Iris segmentation is to separate the iris area from the entire iris image,so as to carry out iris recognition research.The traditional iris segmentation algorithm cannot accurately segment images blocked by light spots,eyelashes,and eyelids.The segmentation effect of glasses blocking images is even more unsatisfactory.Convolutional neural networks have shown very high segmentation performance in image segmentation,but they cannot achieve good segmentation results directly in iris segmentation tasks,and a feature fusion network based on attention mechanism(Attention mechanism Feature Fusion Networks,AFFN)is proposed for the problem.The specific work is:(1)Because the ground-truth of the iris image is difficult to label,there are few samples in the current data set,and the iris data set has been expanded by means of brightness change,rotation,cropping and so on.(2)In order to solve the problem of information loss caused by the pooling layer of the fully connected network(Fully Convolutional Networks,FCN)during image segmentation,a feature fusion method was designed by analyzing the feature maps generated by different convolutional layers of FCN.(3)Add weights to the feature channels after fusion,strengthen effective features,and discard invalid information.(4)Experiments on three iris datasets of CASIA-Iris-Thousand,CASIA-Iris-Interval and UBIRIS,compared with other algorithms,the segmentation algorithm proposed in this paper is very robust and achieves accurate iris segmentation.In this paper,the iris recognition on the open set is studied,and the categories of the test set and the training set are not exactly the same.Traditional feature extraction algorithms are less robust,and deep learning-based methods mostly regard iris recognition as a classification task,which is not suitable for the research of open set iris recognition.In view of the above problems,a siamese network based on residuals(Siamese Network of Residual,Res-Siamese)is proposed.The specific work is:(1)Expand the iris image through changes in brightness and angle to study the robustness of feature extraction algorithms.(2)Improve the residual module of the residual network,increase the feature extraction capability,and remove the batch normalization layer to reduce memory consumption.The improved residual block is used in the twin network for special extraction.(3)In the CASIA-IrisInterval data set,an unnormalized iris image is used for experiments.The statistical diagram of the Euclidean distance between features shows the accuracy of the features.The comparison experiment proves that the iris recognition algorithm designed this article has a high accuracy.