Research On Contactless Palm Vein Recognition Technology Based On Deep Learning

With the rapid development of the current economy and information technology,how to effectively improve the security of citizenship information is particularly important.The palm vein recognition technology is a biometric recognition technology with high safety and high anti-counterfeiting,and will have a broad application space in the future.However,most of the current palm vein recognition algorithms are based on artificially designed feature for recognition.These features are not robust and sensitive to image quality and palm posture.Moreover,the current research on palm vein recognition is generally based on images which acquired from contact devices.And there were few studies on contactless palm vein recognition.The palm vein image which collected from contactless devices often has large deformation,such as palm bending,palm tilt or palm over stretching,and the illumination is not uniform enough,so it has higher recognition difficulty.However,compared with the contact type,the contactless collection method has a better user experience,and its application prospect is broader.In view of the above problems,this paper adopts the deep learning method to research the problem of contactless palm image landmark localization and feature extraction,and has achieved good improvement effect.The main research contents of this paper are as follows:1)For the problem that the traditional algorithm is difficult to accurately locate the landmark of the non-contact palm image,this paper proposes a deep learning method to locate the landmark of the palm,and designs a two-level cascade convolutional neural network to locate the landmark of the palm;In order to improve the accuracy of the landmark localization,we improved the MSE loss function and proposed Modified MSE Loss function for training,which effectively improves the landmark localization accuracy of the palm;2)Owing to the contactless palm vein database has fewer samples,training neural network is prone to over-fitting.We proposed intra-class data augmentation and inter-class data augmentation method,which effectively expands the original contactless palm database and provide support for training the palm vein feature extraction network;3)After analyzing the design principle of the classical classification neural network structure,we designed two deep feature extraction networks which named Resnet18-Modified and VGG16-Modified,then improved the Resnet Residual Block module structure.The experimental results prove that the improved Residual Block module can effectively improve the recognition accuracy.In addition,we also explored the influence of different feature vector dimension on accuracy.Experiments show that the 512-dimension feature vectors achieve the highest accuracy on large deep network,while 256-dimension compact feature vectors have higher precision on lightweight network;4)In view of the problem that the traditional deep neural network has large parameters and high calculation consume,this paper designs a lightweight network named TinyPVNet,which mainly uses deep separable convolution structure,which effectively reduce network parameters and model inference time.we achieved 0.51%equal error rate in the CASIA dataset.In addition,we effectively compresses the TinyPVNet model volume by quantifying weight,and enabling the deployment of lightweight networks on platforms with limited computing resources such as embedded platform.