The Research On Deep Learning Based Finger Vein Biometric Recognition System

With the development of informatics technology,people pay more and more attention to the protection of personal information,so reasearcher need a reliable and convenient identification technology.As the development of computer science,biometrics recognition technology is widely concerned.The first generation of biometrics recognition technology like: face recognition technology,fingerprint recognition and iris recognition have been widely used in many fields.As a new biometric recognition technology,finger vein recognition has unique advantages like outstanding safety and vitality,which makes it has very high development prospects and value.However,there are some problems needed to be solved such as poor image quality and randomly placement of finger,which limits the application of finger vein recognition technology.In order to improve the performance of finger vein recognition system,we make some improvements about it.The main contributions of this article are as follows:First of all,for the problem of extracting the region of interest of finger vein image in complex background,a new semantic segmentation neural network based algorithm is proposed,while combining with the estimation of the rotation angle of finger and the position of finger joint,which helps the system to perform geometric correction for image of finger vein,and effectively reducing the impact caused by the difference among different user.Second,for the problem of low quality of images which make great impact on the finger vein recognition system.By analyzing the features of finger vein recognition,using image processing algorithm to preprocess the image and enhance the texture,and then by threshold and morphological processing,pseudo texture will be removed,the area of finger vein texture is used to estimate the image quality of image.It helps us to reduce the effect of image quality on the performance of the recognition system.Thridly,for the problem of insufficient feature extraction ability of traditional algorithm,a deep learning based vein feature extraction algorithm is proposed.Choosing the loss function by analyse the characteristics of finger vein recognition task.Meanwhile,for the design of basic network,following improvement are proposed,change the pooling algorithm from feature map to feature vector,light weight network design and improvement of reusing feature map.It helps the network's performance maintain the same,meanwhile the network only has 2% parameter of normal network and 2.5 times faster than normal network.Finally,for the problem that during the 1:N recognition the time increase linearly with the increasement of template feature vectors.Three different algorithms are compared in nearest beighbor search field,and choose the precise search algorithm based on space segmentation,and then improve this algorithm to achieve faster speed than linear search in high dimension.In this thesis,the important parts of the finger vein recognition system are improved,and the effectiveness of the improvement is verified through comparative experiments,and finnaly a finger vein biometric recognition system is realized.