Research Of Finger Vein Recognition Algorithm Based On Deep Learning

Finger vein recognition is a kind of identification technology that obtains finger vein patterns under near-infrared light as a biometric feature.Compared with the traditional biometric technology,this method has significant advantages.Traditional finger vein recognition technology mostly relies on hand-made design features based on various mathematical hypotheses,and experience definition and human intervention will bring errors.The depth feature learned from convolution neural network has good generalization and expression ability,but it is limited to modeling larger and more complex vein features,and does not consider the spatial dependence of finger vein pixels.In view of the above problems,this paper has done the following works:1.An improved ROI extraction method and label making method are proposed.Aiming at the phenomenon of missing pixels in some data sets,in the routine ROI extraction process,this paper adds the method of sliding window to find the sum of pixels for filtering,and uses the self-built tool to adjust manually.This method has certain universality and is robust to the image of vein missing.In view of the current label making cannot make good use of the advantages of benchmark segmentation,this paper gives different weights to the traditional baseline to allocate labels.This method is more accurate than the simple fusion label and combination label,and more in line with the actual application.2.A segmentation algorithm of finger vein based on fully convolutional neural network and conditional random field is proposed.In view of the current convolution neural network lack of internal mechanism to deal with fixed geometric transformation and smooth constraints to encourage the consistency of edge,appearance and space between adjacent pixels,this paper synthesizes the advantages of recurrent neural network,residual neural network,deformable convolution network and conditional random field,and makes convolution operation by adding offset to the sampling grid in normal convolution According to the size and shape of the vein,the receptive domain can be adjusted adaptively,so that the vein of various shapes and sizes can be captured.These more complex and deeper features are mined and accumulated by the recurrent neural network and residual neural network.Conditional random field is introduced to refine the output of the fully convolutional neural network,which considers the label consistency among similar pixels,making the segmentation result more accurate.Conditional random field can be inferred as the embedding of the whole network model,and conventional back propagation algorithm is used for end-to-end training.The experimental results show that the fully convolutional neural network proposed in this paper is better than other equivalent models in segmentation effect,and the embedded conditional random field also further improves the performance of the system.3.A finger vein verification algorithm based on cascade optimization IU-Net is proposed.In view of the current problems such as the small datasets of public finger vein,the loss of structural correlation in network feature expression,and the difficulty in weighing the relationship between network depth,width and computing resources,this paper takes the contraction path of U-Net network as the infrastructure,and the repeated structure of network composed of conventional convolution and residual recurrent convolution is used to extract deeper features.By introducing the Inception module and its variants into the middle layer and back layer of the network,we can deepen the depth of the network,increase the width of the network by extracting multi-scale information,and save a lot of computing resources.In the process of network training,the difference images of two images are used as input to augment the data.The proposed adaptive weight cross entropy and gradient harmonized cross entropy can be reasonably distributed according to the number of outliers in the data set,so that the loss weight of each training process can be dynamically adjusted,and the problem of category imbalance can be effectively solved on the premise of making full use of limited data resources.Finally,the difference image and the original segmented image are used to connect the channels,and the pre-training network is fine tuned to achieve cascade optimization.The experimental results show that this method has the lowest equal error rate at present.