Finger Vein Recognition Based On Multi-scale LBP And Mproved Deep Confidence Network

In modern society,due to the importance of security on individual,information,and property,personal identification has received widespread attention.In recent years,finger vein recognition has become one of the research hotspots of personal identification due to its unique advantages.Compared with other identification method,the advantages of finger vein recognition are reflected in living body recognition,implicit features,ease of use,and high level of security.The basic process of finger vein recognition includes image acquisition,preprocessing,feature extraction and feature recognition.Among them,feature extraction and feature recognition are the key procedures of finger vein recognition.Extracting more accurate and comprehensive information from images and using this information more efficiently have always been the focus and difficulty of vein recognition.This article studies these two key procedures in depth,and the research is as follows:(1)LBP feature extraction based on multi-scale fusion: To solve the problem that the traditional LBP feature extraction can only extract histogram features at a single scale,this paper presents a multi-scale fusion LBP feature extraction that uses LBP feature histograms at multiple scales.The method of fusion extracts the histogram features of the finger vein image.(2)Improved Deep Belief Network: This paper proposes an improved Deep Belief Network for feature classification.It learns LBP histogram features and builds models to complete finger vein recognition.In the improved Deep Belief Network,for the problem of long training time of Deep Belief Network,this paper successfully reduces the training time of Deep Belief Network by introducing an adaptive learning rate mothed and using Leaky ReLU as the activation function.(3)Based on the combination of the above methods,this paper proposes a finger vein recognition method based on multi-scale fusion LBP feature extraction and Deep Belief Network.After experiments,this method has a training time of 247.5s and a 99.70 accuracy of 99.70%.