Research Of Palm Vein And Thenar Palm Vein Based On Deep Learning

The palm vein(hereinafter referred to as "palm vein"),including the palm and the thenar palm vein are hidden features,located within the skin of the palm,with a complex structure and difficult to be copied.Palm vein recognition has a high degree of uniqueness,security and stability.It is more stable than face recognition,safer than palm print and fingerprint recognition,and easier than iris recognition capture.At the same time,it can be used as the basis for living body identification.With the development of deep learning technology,palm vein image feature recognition based on deep learning has good research and application value.Palm vein recognition mainly contains three stages of image pre-processing,feature extraction and classification recognition.Currently,the palm vein has a low accuracy rate compared to explicit biometric features such as palm print and fingerprint.This paper uses the deep learning method to study the multimodal fusion recognition of two features through multiple fusion methods,based on the study of two unimodal recognition improvement methods for palm vein and thenar palm vein.At the same time,study the palm vein recognition based on the knowledge distillation in order to apply fast recognition application scenarios.The main research work of this paper is as follows:(1)Research on techniques related to palm vein image pre-processing.Firstly,this paper used palm vein discrimination model based on the FDR to determine the 850nm spectral palm vein database as the main dataset for the subject research,then the palm vein images were preprocessed.Specifically includes:palm vein image binarisation,morphological processing,contour extraction,key point localization,palm vein and thenar palm vein ROI extraction,normalisation,image denoising,image enhancement,palm vein pattern segmentation,aim vein pattern refinement and other processing.Among them,for the problem of hand key point localization low accuracy in traditional method,this paper adopted a convolutional neural network localization method for hand key point localization,which effectively improves the localization accuracy.In addition,designed an improved ROI extraction method for the thenar palm vein,which can perform the thenar palm vein ROI extraction more accurately.(2)For the problem of DenseNet network low recognition rate in palm vein recognition,this paper added residual calculation to the dense network,designed an improved palm vein recognition method based on DenseNet network,and investigated the effects of learning rate,neurons numbers in fully connected layers and optimisation algorithm on recognition results to obtain the optimal network structure and parameters.In addition,proposed four solutions for the overfitting:rotation angle expansion dataset,L2 parameter regularization,Dropout and batch normalization,so that the recognition rate of palm vein and thenar palm vein can reach 97.60%and 98.40%respectively,effectively improved the palm vein recognition accuracy.(3)Studied the palm vein and thenar palm vein fusion recognition method.Based on the unimodal classification recognition results of palm vein and thenar palm vein in step(2)and the optimised network model,respectively designed a weighted fusion recognition model and an improved DenseNet+SVM feature fusion recognition model of palm vein and thenar palm vein.The experimental results show that the two fusion recognition methods respectively improved the recognition accuracy to 99.2%and 99.9%.(4)Studied the palm vein recognition method based on knowledge distillation and attention mechanisms.For the problem that the optimized network model in step(2)has many parameters and is bulky,designed a more efficient palm vein recognition method based on knowledge distillation model in order to save running time and be suitable for the fast palm vein recognition application scenario.The VGG-19 network was used as the teacher network,and a lighter network with only five convolutional layers and three fully connected layers was selected as the student network,while added an attention mechanism to the student network to enhance the student network’s ability to learn complex features,and used the student network to classify and recognize palm vein.Experiments on the CASIA-M database show that although the teacher network recognition rate is 9.52%higher than the student network,the teacher network model is about four times larger than the student network and the recognition time is about three times longer than student network,so using the student network for recognition can save more storage space and running time.