Font Size: a A A

Palm Vein Recognition Research Based On Adaptive Fusion Image Enhancement And Deep Learning

Posted on:2021-05-13Degree:MasterType:Thesis
Country:ChinaCandidate:M Y LouFull Text:PDF
GTID:2428330605958362Subject:Biomedical engineering
Abstract/Summary:PDF Full Text Request
With the coming of the 21 st century,people's security awareness has been greatly improved,and the traditional identification methods are difficult to meet the needs of the current society.Biometric(including dominant and recessive characteristics)identification,which is based on human physiological or behavioral characteristics,has better universality,uniqueness and stability.Compared with the dominant biological characteristics,the recessive biological characteristics are more secure and reliable because they can't be imitated or forged.The palm vein is a biological characteristic in vivo,which can't be forged or copy,and it belongs to the recessive characteristics.Palm vein recognition is more secure than palmprint,fingerprint,face and other dominant characteristics,more stable than voice and gait,and easier to collect than iris.Therefore,palm vein recognition has gradually become the focus in this field.Palm vein recognition is safe and controllable,but its intradermal characteristics make the quality of acquisition image low.It is necessary to adopt more effective image enhancement in the preprocessing stage and more advanced technology for image recognition.For palm vein recognition in this paper,a palm vein image enhancement method based on adaptive fusion is proposed,and an advanced palm vein recognition method based on deeep learning is proposed simultaneously.First,image enhancement method based on adaptive fusion:according to the image evaluation index,the advantages and disadvantages of different image enhancement methods are determined,and the fusion weights of different enhancement images are set in an adaptive way to achieve the complementary effect of different enhancement methods.For the palm vein image,as an example of adaptive fusion image enhancement,this paper lists two different adaptive image enhancement methods and adopts two different adaptive fusion strategies.(1)Palm vein image enhancement method based on POSHE and DCP fused adaptively.Firstly,according to the variation coefficient of the image,the defogging coefficient of Dark Channel Prior(DCP)defogging algorithm is selected adaptively.Partial block overlapping histogram equalization(POSHE)algorithm and adaptive DCP algorithm are used to obtain their enhanced images respectively.Secondly,the image is divided into 16 sub-blocks,and the adaptive fusion weight of each sub-block was determined by the mean and the standard deviation.Finally,two kinds of enhanced images are fused adaptively according to the weight.The correct recognition rates of PolyU,CASIA and self-built databases are 99.98%,94.27%and 92.05%,respectively.(2)Palm vein image enhancement method based on adaptive fusion and Gabor filter.Firstly,the POSHE enhanced image is obtained based on POSHE algorithm,and the difference of Gaussian(DOG)enhanced image is obtained based on DOG algorithm.Then,two kinds of enhanced images are fused adaptively according to the second-order statistics of each gray-level co-occurrence matrix.Lastly,the multi-directional Gabor filter is used for enhancement.The correct recognition rates of PolyU,CASIA and self-built databases are 99.96%,94.50%and 94.89%,respectively.Second,palm vein recognition method based on deep learning:it achieves better recognition effect by optimizing and improving the existing neural network model.Firstly,for the enhanced images with significantly improved quality,the ResNet model with strong generalization ability is used for feature extraction,and the residual block can effectively alleviate the network degradation.Then,optimize and improve the ResNet network structure by using ELU to replace the original Relu activation function and adding dropout layer.It can alleviate the gradient disappearance,prevent over fitting and accelerate the convergence of the model.Finally,combine with the idea of dense network.The original information is input into the multi-level convolution layer to enhance the richness and effectiveness of the extracted features.Experiments are carried out on PolyU,CASIA and self-built databases respectively,the results show that the recognition method based on deep learning in this paper is more effective.
Keywords/Search Tags:Palm vein recognition, Image enhancement, Adaptive fusion, Deep learning, ResNet
PDF Full Text Request
Related items