Research On Denoising And Recognition Of Finger Vein Image Based On Deep Learning

As people attach great importance to personal information security,finger vein recognition technology has been applied to identity verification in daily life due to its strong security and convenience advantages.However,due to the special application scenarios of the finger vein recognition system,the vein images are easily affected by factors such as mirror dust,mirror stains,and finger stripes during the collection process,resulting in the collected vein images carrying noise,which seriously damages the texture information of the veins,so it greatly affects the subsequent recognition performance.This thesis firstly studies and analyzes the existing denoising algorithms.It is found that they are all denoising based on images with clear texture.When denoising the vein image with insignificant contrast between vein texture and background area,the edge information of vein image is not fully retained and vein texture is not restored.Therefore,a finger vein denoising algorithm based on the gradient direction residual structure combined with LBP(Local Binary Pattern)texture constraints is proposed,the edge and texture information of vein image are effectively preserved while denoising is realized;Secondly,a lightweight digital vein recognition algorithm based on deep residual network is proposed to solve the problem that the number of network parameters is too large to be deployed to mobile devices,which combines the characteristics of finger veins to design a lightweight network,and uses knowledge distillation restores its performance,which ensures the recognition rate of fingers in normal and special postures while implementing the compression network.It has a good reference value for denoising and recognition research on this type of image.The specific research contents are as follows:1.A finger vein denoising algorithm based on gradient directional residual structure combined with LBP texture constraints is proposed.First of all,gradient direction operator is introduced into residual structure to solve the problem of losing gradient direction information in de-noised vein image.Then,the local features such as texture in the shallow layer of the image are fused with the semantic information in the deep layer through the skip layer connection,so as to retain the global and local information of the denoised vein image as much as possible.Finally,the LBP texture loss term is added to the loss function to enhance the recovery ability of vein texture.Through theoretical analysis and experiments,it is proved that compared with the traditional three-dimensional block matching(Block Matching 3D,BM3D)denoising algorithm and the deep learning denoising algorithm based on Dn CNN(Denoising Convolutional Neural Network)and FFDNet(Fast and Flexible Denoising Network)network,the finger vein denoising algorithm based on the gradient direction residual structure combined with LBP texture constraints achieves denoising and has more excellent performance in visual effects,peak signal to noise ratio(Peak Signal to Noise Ratio,PSNR)and recognition performance.2.A lightweight finger vein recognition algorithm based on deep residual network is proposed.First,use depthwise separable convolution instead of convolution to build a network residual block,and add SE(Squeeze and Excitation)attention mechanism module to the residual block to extract detailed features in the spatial domain of finger veins,and on this basis introduce Width scaling factor to further compress the network model.Second,the network is jointly supervised using knowledge distillation loss,Curricular Face and cross-entropy loss,which solves the performance degradation problem of lightweight deep residual network due to the small amount of parameters.Simulation experiments show that when the fingers are placed in normal and special postures,compared with the widely used lightweight network model Mobile Face Net,the lightweight finger vein recognition algorithm based on the deep residual network can still maintain a better recognition effect.The recognition performance is ensured while realizing the lightweight of the network.