Research On Palmprint Recognition Technologies Based On Deep Learning

With the rapid development of Internet technologies as well as the prevalence of mobile terminals,stringent security,practicability and effectiveness of access control systems are required.Palmprint recognition technology provides new train of thought for the deployment of more secure and reliable access control systems,strengthening private information protection and achieving high-performance identity recognition.Also,it has become a research focus in the fields of pattern recognition,computer vision and information security,and has been widely used in the areas of people's livelihood,finance,safe-protection,criminal investigation and counter-terrorism.In recent years,the application scenes of palmprint recognition have changed from highly constrained contact system to mobile terminal and Internet application,which brings a series of problems of low-constrained palmprint recognition.Since image acquisition does not limit the acquisition environment and hand placement,there are widespread problems in palmprint images such as rotation,scale change,motion blur,noise,poor light conditions and complex background.Therefore,it is very challenging to perform low constrained palmprint recognition.To solve this problem,this dissertation proposes palmprint recognition methods based on deep learning,which focuses on low-constrained online palmprint recognition and low-constrained police palmprint recognition.(1)A palmprint region of interest(ROI)extraction method under low constraints is proposed,which is able to deal with the complex environment such as light change and hand movement in palmprint data collection.Principal Component Analysis and Pearson's Bimodality Index are used to find out the most discriminative color component candidate for hand segmentation,which aims to separate light influence and reserve the inherent chrominance of an palmprint image.Learning-based Partial Least Squares Regression algorithm is proposed to obtain regression coefficient between original image and ground truth,which aims to record hand shape in an image.In addition,a boundary tracking algorithm based on topological analysis is proposed to traverse the outer contour of all the blocks in the binary image,and the threshold is set to remove the interference items and retain the contour of the hand.Hand geometry-driven convex hull searching algorithm is proposed to detect valley points between fingers(i.e.the key reference points),and determine the location of palm ROI region.The experimental results show that the proposed method achieves preferable ROI extraction accuracy.(2)An online palmprint recognition method based on gender priori knowledge is proposed,which makes use of the soft biometrics trait of gender to supplement the identity information provided by palmprint,and improves the performance of online palmprint recognition with low constraints.The proposed method innovatively incorporates gender feature into the decision-making process of palmprint recognition,fully considers the complementarity between gender feature and identity information,and is conducive to enhancing the discrimination of palmprint feature.First,two deep learning models are proposed to study identity recognition and gender classification sub-problems respectively.Then,two fusion models of different architectures are proposed to explore the boosting effect of different feature fusion ways on palmprint recognition.In the process of training,the fusion model of parallel architecture can extract identity information and gender feature synchronously.In the fusion model of sequential architecture,gender feature and identity information are sequentially extracted through two-stage model training,and palmprint recognition is completed in the second stage.The experimental results show that the proposed method is able to achieve higher palmprint recognition accuracy.Besides,the gender priori is able to improve the accuracy of palmprint recognition and enhance the confidence of palmprint recognition.(3)A robust contactless palmprint recognition method is proposed to realize high accuracy palmprint recognition under blurry and noisy environment.A palmprint recognition model based on residual network is designed to retain more image features and pass them downward by using skip-connection.The designed model is able to effectively alleviate the problem of gradient disappearance in network training.In order to explore the anti-blur and anti-noise performance of the proposed method,degraded palmprint images are generated by adding multilevel motion blur,Gaussian noise and salt&pepper noise into the original ROI images,and the correct identification rate and equal error rate are calculated again based on the corrupted data.The experimental results show that the proposed method is able to achieve higher palmprint recognition accuracy as well as the strong robustness in palmprint recognition on degraded palmprint images.(4)An end-to-end palmprint recognition method in open environment is proposed to effectively cope with image deformation,distortion,warp and other complex environments by locating and aligning ROI of nonlinear transform.To solve the palmprint recognition problem based on evidence images,this dissertation proposes a deep learning model containing multi-module convolutional neural network,which consists of ROI localization and alignment network,feature extraction and recognition network,and an in-network data augmentation based on dropout.This method uses the whole hand image instead of palmprint ROI as the input to complete ROI extraction,palmprint feature extraction and matching in an end-to-end fashion.The experimental results show that this method is able to effectively solve the problem of police palmprint recognition in criminal investigation.