Research On Palm Vein Feature Segmentation And Recognition Algorithm Based On Deep Adversarial Learning

Palm vein recognition has become one of the key research directions in the field of biometric identification because of its advantages such as uniqueness,in vivo recognition and non-contact nature.In recent years,deep learning with powerful feature representation has shown good performance in palm vein segmentation and recognition.However,deep learning models require a high amount of training data.In practical applications,the small number of registered samples per person in the dataset due to privacy protection leads to the limited performance of deep learning models in palm vein segmentation and recognition tasks.How to effectively segment and recognize palm vein features with limited vein data samples is a pressing problem in this paper,and deep adversarial learning can be used for data augmentation and improving model generalization,so this paper applies deep adversarial learning to palm vein feature segmentation and recognition,and conducts the following specific research:(1)A palm vein feature segmentation algorithm based on deep adversarial learning is proposed.First,a convolutional neural network-based model is established to transform the palm vein images and gold standard(Ground Truth)to produce vein images and gold standard with diversity.Then,a vein segmentation model based on U-Net network is constructed.Finally,the transformation model is combined with the segmentation model to build a vein pattern segmentation model based on adversarial learning.The convolutional neural network-based transform model aims to generate challenging sample pairs(vein images and gold standard)to increase the segmentation difficulty of the vein segmentation model,while the vein segmentation model can learn more robust feature expressions using challenging sample pairs to improve its generalization ability.To prevent semantic collapse of vein images,cosine similarity is added to regularize the transformed sample pairs.Experimental results on three public palm vein datasets,CASIA,VERA and Poly U,show that the deep adversarial learning-based palm vein feature segmentation algorithm outperforms existing vein segmentation methods with equal error rate(EER)of 0.33%,0.59% and 0.6%,respectively.(2)A deep adversarial learning based algorithm for palm vein recognition is proposed.Firstly,conditional generative adversarial networks(c DCGAN)are used to learn the true distribution of the data.The trained generator can map the hidden variables to the sample space.Then,the trained generator and vein classifier are combined to build a vein recognition model based on adversarial learning,which uses adversarial learning to update the input set(set of hidden variables)of the generator and the classifier alternatively.During training,the model increases the training loss of the vein classifier by learning a set of hidden variables and generating challenging vein samples,while the vein classifier is able to learn more robust vein features using the challenging samples to improve the recognition performance.To generate high-quality vein samples,the exponential sliding average model and cosine distance are introduced as regular terms into the loss function for training the model.Experimental results on Poly U,Tongji University and VERA palm vein datasets show that the method in this thesis outperforms existing data enhancement methods in generating diverse samples and significantly improves the recognition performance of the vein recognition model.