Deep Feature Extraction And Gray Correction Algorithm Research On Finger Vein Recognition

Under the background of the increasing security awareness of people in modern society,biometric identification technology provides an efficient and safe solution to people's identity authentication problem with its superior reliability and convenience.Finger vein recognition,as an emerging biometric technology,has received a lot of attention in research and application fields with its advantages of high security,high stability,and high user affinity.However,there are still some problems that need to be solved urgently in the actual finger vein recognition sys-tem.First,in finger vein recognition,the manual feature extraction algorithm is not robust when dealing with finger rotation offset and environmental brightness changes.Second,neural net-works still have room for improvement in finger vein recognition tasks.In response to the above problems,this article first established a large-scale finger vein data set that includes a variety of actual vein image changes.Secondly,this article introduces deep metric learning method into the feature extraction stage of finger vein authentication,and then use the network to obtain dis-tinguishing features with high generalization and robustness.In addition,this paper designs an efficient attention module to enhance the vein feature extraction capabilities of deep networks.Finally,in order to solve the problem of imaging mode changes under multi-light conditions,this paper proposes a gray-scale correction network.The specific content is as follows:First,this paper established a large-scale finger vein database SCUT LFMB,which aims to solve the problem of low diversity of finger vein images in the public dataset.The data is col-lected by two hardware devices,and the vein images include multi-level illumination conditions,and a wider range of finger rotation gestures,which can make simulation of some problems in the actual authentication scene.Second,in order to solve the problem of insufficient feature extraction capabilities of tra-ditional image algorithms,this paper establishes a vein extraction network based on multi-descriptor aggregation.By adopting metric learning,this network introduce an efficient way to aggregate multiple global descriptors.A single backbone network can be used to complete feature combinations,and training can be completed in an end-to-end mode.There is no need for additional feature diversity control methods.Combining measurement loss and auxiliary classification loss can enable the model to acquire features with strong generalization and dis-crimination.Third,to further improve the ability of the network in vein feature extraction,this paper proposes a joint attention module with both high efficiency and lightness.This module avoids the problem of information loss when using global maximum pooling or global average pooling to compress feature maps.It uses pooling kernal in both vertical and horizontal directions to retain more accurate positional information.In addition,the channel shuffle operation is added to further strengthen the information interaction between the channels.The transformation pro-cess provided by the joint attention only increases few parameter,and is convenient to integrate into a variety of commonly used model structures.Fourth,to deal with the problem of changes in the lighting conditions of the vein image in the actual test environment,this paper designs a U-shaped gray-scale correction network based on the assumption of image dehazing.The network can perform adaptive parameter estimation on the input vein image imaged under different lighting conditions,and directly output the cor-rected image,so as to realize the effective adjustment of the gray distribution of the input vein image.Experiments prove that the generated images of the network have good imaging quality and can improve the robustness of the system during the authentication process.This paper has completed multiple sets of comparative experiments on multiple public databases and large self-built databases.First of all,the qualitative experimental results in the authentication task verify the effectiveness of the descriptor aggregation network and joint at-tention mechanism proposed in this paper,and EER value of 0.314%,0.067%,0.113% has been achieved on the three vein databases SDUMLA-HMT,MMCBNU?6000 and FV?USM,respec-tively.Secondly,the image generation quality and recognition performance analysis experiment proves the positive role of the gray correction network in the recognition task.