Study Of Finger Vein ROI Extraction And Recognition Based On Deep Learning

In recent years,with the popularity of smart devices using such as fingerprints and faces as the verification method,biometric identification technology has entered people's attention with its advantages such as quickness,convenience and high confidentiality,and has been developing rapidly in recent years.Finger vein recognition is a new type of biometric identification method,which uses the vein distribution pattern formed by subcutaneous superficial veins in the finger area as biometric features.When collecting the vein data,the finger needs to be placed in an infrared collection device,and since the finger vein is located inside the human body,it is difficult to be stolen and there is almost no risk of forgery.In addition,finger vein recognition has a lower false rejection rate and better stability than other more traditional biometric technologies.Therefore,in recent years,the research interest on finger vein recognition has been gradually increasing in related fields.Generally,the process of finger vein recognition includes ROI(Region of Interest)extraction,feature extraction,feature matching,etc.Since the convolutional neural network was proposed,deep learning has improved the performance of image recognition compared with traditional methods,and is now widely used in biometric recognition.In this paper,we improve finger vein recognition performance by solve problems use deep learning methods in some process.The work and results of this paper are as follows.1.This paper proposed a deep learning-based ROI extraction method.The traditional rule-based method is highly sensitive to background noise,image size,and image gray value,and has low robustness.In this paper,a U-Net-based ROI extraction method is designed and implemented to perform finger semantic segmentation using a U-shaped network structure,which is able to retain the complete finger contours.It is experimentally verified that this scheme has strong generalization ability and robustness,and can still extract ROI regions more accurately in the presence of noise interference.2.This paper proposed a GAN(Generative Adversarial Networks)based method for finger vein data generation.Convolutional neural networks require a large amount of labeled image data for training in order to obtain a model with good performance.However,there are few public datasets related to finger veins,which make it difficult to adequately train complex neural network models.In this paper,a finger vein data generation method is proposed,in which the finger vein pattern are generated using a simulation growth algorithm after analyzing the finger vein anatomy,and then convert the finger vein pattern into finger vein images using conditional generative neural network,and a total of 5363 finger vein images of 5363 classes are generated,and the validity of the generated data is verified by experiments.3.This paper proposed a ResNet-based finger vein spatial attention recognition network.The Shortcut structure allows the convolutional neural network model to be deepened continuously without the performance degradation caused by excessive model complexity.However,deep models require a large amount of machine resources and are difficult to train.In this paper,the network structure based on ResNet is improved by introducing a spatial attention mechanism to enable the model to perform feature selection in the channel dimension.The network adopts modular design,which can be easily transplanted to the existing network.After experimental comparison,this method can achieve better recognition performance with a small increase in the number of parameters compared with ResNet of the same depth in finger vein recognition task.