Research On Texture Image Feature Extraction And Classification Algorithm Based On Finger Vein Recognition

With the further development of finger vein recognition algorithm in recent years,as a representative of the second generation of biometric identification technology,finger vein recognition is widely used in various mobile devices due to its advantages such as high accuracy,unforgeability and anti-interference.Therefore,the study of lightweight algorithms applied to this technology has become particularly important.To improve the extraction efficiency of the feature extraction algorithm and the performance of the recognition algorithm,this paper presents an in-depth study of the traditional feature extraction algorithm and a lightweight network model.To suppress image noise and better extract finger vein main pattern information,a feature extraction algorithm based on LBP(Local Binary Pattern)is proposed in this paper.Firstly,the concept of local pixel dispersion is proposed to extract the local pixel dispersion of the original image to process the image for noise reduction and highlighting the main stripe information;Secondly,a new way of obtaining binary sequences and weights is proposed,using the values of pixel dispersion distribution maps as the reference for obtaining binary sequences and the order of dispersion as the basis for calculating weights,which provides a better basis for weighted calculation of feature information.Experimental validation was conducted on three publicly available finger vein datasets,and the proposed algorithm was experimentally proven to have excellent performance in all metrics,excellent performance in ROC curve,and its performance is better than other algorithms,and the time complexity is reduced by 20% compared to LDP,which has better advantages in comprehensive performance.A dual-channel based probabilistic fusion model is proposed to address the need for lightweight recognition algorithms for finger vein mobile devices and to further improve the comprehensive performance of deep learning algorithms.The purpose of the model is to lighten the algorithm while ensuring its recognition accuracy.Firstly,in terms of network structure,a method of separating horizontal pooling and vertical pooling is designed to reduce the weakening effect of the pooling layer on neurons,which can maximize the retention of neuronal information;Secondly,trainable bias is proposed to better obtain the conditional probability by using trainable bias instead of hidden layer of activation function in a single channel.On the one hand the method reduces the nonlinear mapping of the activation function and provides a better basis for probabilistic fusion;On the other hand the method compensates the feature information,which improves the recognition accuracy of the network to some extent;Finally,a Bayesian fusion strategy with non-average distribution of prior probabilities is proposed,and the overall network model adopts a dual-channel model to improve the accuracy of output probabilities through probability mapping.Experimental validation was performed on three publicly available finger vein datasets,FV＿USM,SDUMLA and UTFVP.With only 5.57 M training parameters and without any preprocessing,the algorithm in this paper achieves an average correct recognition rate of more than 92% on the three datasets,and the recognition performance is better than the lightweight algorithms such as Mobile Net and Shuffle Net,with excellent comprehensive performance,which provides a reference for the algorithm selection of mobile devices.