Research On Deep Feature Learning And Recognition For Hand-vein Image

Hand-vein recognition technology has a very broad application prospect in biometric identification,and is a research hotspot in the field of biometric recognition.Hand-vein recognition systems include several main steps such as vein image acquisition,vein image preprocessing,feature extraction and feature matching.As one of the most important steps of vein recognition systems,the effect of feature extraction algorithms for vein images directly affects the recognition rate and robustness of vein recognition systems.In recent years,Deep Convolution Neural Networks(DCNN)have been widely used in large-scale image recognition tasks,and have achieved excellent results.However,its representational learning ability depends heavily on the number of training samples.Therefore,it cannot achieve the outstanding peroformance in smallscale image tasks such as vein recognition.At present,deep feature learning for vein images has two basic problems: insufficient training samples of vein images and poor ability of representation learningof network models.Therefore,to solve the aboved problems,this thesis carries out the researches on deep feature learning of hand-vein images for improving the accuracy and robustness of hand-vein image recognition algorithms,and constructs a series of more discriminative and robust hand-vein recognition algorithms.The main contributions and innovations of this thesis are as follows:1.The construction of hand-vein database.To solve the problem that there is insufficient public hand-vein image data at present,this thesis constructs a hand-vein image database.Firstly,the acquisition device of vein images is designed to obtain the dorsal hand vein images and palm vein images.Secondly,the preprocessing methods of hand-vein image is utilized to extract the regions of interest(ROI)of hand-vein images.Finally,a large-scale hand-vein image database which contains 5720 images is constructed in this thesis.2.Multi-layer convolutional features fusion for hand-vein recognition.To solve the issue that DCNN cannot better obtain more discriminative ability for vein recognition due to the insufficient training samples,a vein recognition method based on a pre-trained convolutional neural network is firstly proposed in this thesis.In this method,the response characteristics of convolutional activations based on vein information are firstly analyzed,and a local max-pooling method of preserving spatial position information for high-order convolution feature map is designed.Then,a fusion model of multi-layer deep convolution features is constructed,which makes full use of the detail information in low-level convolutional features and the semantic information in high-level convolution feature map,to improve the deep representation ability of the proposed network model and increase the accuracy of vein recognition methods.3.Multi-scale deep representation aggregation for hand-vein recognition.To solve the issue of insufficient removal of non-vein and noise information from convolutional activations in vein recognition methods based on a pre-trained DCNN,a multi-scale deep representation aggregation method based on a hierarchical feature selection model is proposed in this thesis.This method deeply analyzes the response characteristics of convolutional activations with vein information,and reveals the learning mechanism of high-level semantic information of DCNN based on vein information.A hierarchical deep feature selection model is constructed,which effectively removes the non-vein and noise information contained in the deep convolutional features,further improves the representation ability of deep representations and increases the recognition rate of our proposed hand-vein recognition model.4.Disentangled representation network for hand-vein recognition.To solve the problem of insufficient feature representation and poor robustness of single vein texture feature extraction algorithms and vein shape feature extraction algorithms,a disentangled representation network based on multi-scale attention residual module is proposed in this thesis.In this method,vein texture feature coding network and shape feature coding network are constructed to realize the adaptive decoupling of texture and shape features of vein images.A weight-guided highly discriminative deep feature learning module is designed to reveal the influence mechanism of vein texture features and shape features for effectiveness of vein recognition models,which enhances the representation ability of deep vein features,and then improves the performance of handvein recognition algorithms.5.Hand-vein recognition based on synthetic data.Aiming at the problem that the DCNN has insufficient representation ability for real vein images due to the domain shifts between generated vein samples and real vein samples,a deep feature learning model based on synthetic vein samples is proposed in this thesis.This method constructs a vein generation model based on disentangled representation learning,which improves the quality of vein generation samples.An adaptive fusion network for vein images is designed to reduce the domain gaps between the generated vein samples and the real vein samples.A global-local deep vein feature learning module is proposed to further enhance the ability of feature representation of DCNN for vein images and improve the recognition rate of our proposed hand-vein recognition models.This thesis includes 62 pictures,45 tables and 139 references.