Research On Finger Vein Recognition Technology Of Deep Learning

Finger vein recognition technology is a new biological feature recognition technology.Its medical basis is that human blood can absorb specific wavelength of light,so finger vein can be clearly imaged by using specific wavelength of light to irradiate the finger.At present,finger vein recognition technology has entered a mature stage,but there are still some shortcomings in practical application.Common instruments will decrease the recognition speed with the increase of matching number,or when the ambient temperature is low or high,the reaction speed will also slow down.In order to solve this problem,we can only break the routine,instead of using hardware facilities,we can enter the server operation through the network.In this way,it will no longer be limited by the hardware memory capacity,nor will it be subject to changes in the ambient temperature leading to slow device response.Various network-based finger vein classification algorithms have been proposed at home and abroad based on deep learning,but the common CNN has the shortcomings of information loss in space.In this paper,an improved finger vein recognition algorithm based on capsule network is proposed.The research of finger vein recognition algorithm in this paper aims at improving finger vein recognition performance.The main research contents are as follows:(1)Aiming at the time-consuming situation of collecting finger vein data sets,an improved algorithm of polar coordinate radial transformation enlarging samples,namely adaptive polar coordinate radial transformation,is proposed.The algorithm transforms the image into polar coordinate image by taking the first non-zero pixel point as zero coordinate,breaking the problem of overlap of transformed images caused by random polar coordinates.In this way,an image can be increased to a multiple of pixels.Experiments on AlexNet with MNIST data set show that the adaptive polar coordinate radial transformation makes the network converge ahead of time,and the accuracy reaches 99.2%.(2)Compared with the classical convolution neural network structure,the accuracy of AlexNet is 71.8%,VGGNet is 88.6%,GoogLeNet is 84.3%,ResNet is89.4%.In view of the important information loss caused by the traditional pooling layer processing of CNN,CapsNets transfers the information from the bottom to thetop in the form of "capsules" in the whole learning process,so that the multi-dimensional features of finger veins are encapsulated in the form of vectors,and the features will be preserved in the network,rather than recovered after the loss.The experimental results show that the network structure characteristics of CapsNets have a better accuracy of 93.7% than that of CNN in dealing with ridge areas.(3)In view of the defect of CapsNets convolution layer,an improved CapsNets,namely convolution capsule network,is proposed.9 * 9 convolution template is not effective in extracting vein vein veins,so 3 * 3 in 4 layers small-scale convolution template is used to extract vein veins edge.The convolution template of 3 *3 in 4layers has doubled the computational parameters and greatly reduced the training time.Experiments show that the convolution capsule network achieves 97.6% accuracy for finger veins and 94.4% accuracy for finger veins database of Hong Kong Polytechnic University.