Research On Finger Vein Recognition And Application Based On Deep Learning

Finger vein belongs to the category of biometric recognition based on physiological features,and recently many excellent algorithms have been produced by the continuous effort of researchers.However,the problems of complex background and low quality of finger vein images obtained by the acquisition device and the poor discriminative feature extracted have been the bottleneck of feature extraction algorithms,so the ability to extract robust and highly discriminative feature becomes the key to finger vein recognition.In order to make the algorithm pay attention to the vein itself while learning,this thesis uses a mask to guide the feature extraction process,and the main research includes two parts: a feature extraction algorithm based on deep learning and a lightweight finger vein segmentation algorithm used to generate the mask.(1)For finger vein recognition,this thesis proposes a feature extraction algorithm based on the open set test protocol,which has achieved good recognition results on three public data sets.In order to eliminate the interference of irrelevant areas,this thesis proposes a "segmentation-assisted classification" idea,that is,using the mask image of finger vein to constrain the feature learning process,so that the network "focuses" on the vein area and gives it greater weight.Specifically,the feature maps of the shallow layer of the network are first sent to the feature pyramid module to fuse primary features of different scales,and then sent to the spatial attention module to obtain the spatial weight map of the image.Based on the results of several classical vein network extraction algorithms,a weighted method is used to obtain a more accurate mask image to constrain the learning of the spatial weight map.Finally,a hybrid loss function combining triplet loss and cross-entropy loss is used to reduce the distance between feature vectors of the same category and increase the distance between feature vectors of different categories in the Euclidean space,thereby improving feature discrimination.The Equal Error Rate(EER)on the SDUMLA,MMCBNU,and FVUSM datasets is 2.50%,0.20%,and 0.14%,respectively.(2)For finger vein segmentation,this thesis designs a lightweight segmentation network with an encoder-decoder structure based on depthwise separable convolution.To address the problems of manual parameter setting and poor segmentation quality when using the traditional vein network extraction method in task one,an idea of using the segmentation results of the deep network to constrain the feature extraction process is proposed.Among them,a multi-scale module and a multiple dilated convolution module are designed in the encoder,which enables the network to capture the difference in vein width while also increasing the receptive field to introduce more context information.The parameter amount of the network is only 0.42 M,and the EER on SDUMLA is reduced by 0.49% through the combination with task one,which proves the feasibility of this scheme.Overall,the model proposed in this thesis has a certain algorithm reference value,and provides a new solution for subsequent finger vein recognition.