Research On Key Technologies Of High-accuracy Touchless Palmprint Recognition

Touchless palmprint recognition technique uses palm images captured at a distance for identity authentication.It has the advantages of rich features,high anti-spoofing ability,flexibility in usage,etc.Therefore,it has recevied more and more attention and has become a hot research topic in the biometrics field.However,flexible palm placement modes and complex acquisition environments will lead to abnormal palm postures and unsatisfactory image quality.On the one hand,variations in palm tilt,rotation,and placement distance will lead to localization errors of the palmprint region of interest(ROI).On the other hand,variations in palm distance and system light source intensities will cause variations in image brightness and lead to information loss.Those factors will adversely affect the accuracy of palmprint recognition.Correspodingly,this thesis mainly studies the palm alignment,pose rectification,ROI localization,and feature coding techniques in touchless acquisition environment.This thesis aims to improve the accuracy of touchless palmprint recognition from the aspects of preprocessing and feature coding.Specifically,the content of this study is summarized as follows:(1)To address the problem of palm misalignment between the visible and infrared modalities caused by the change in palm distance in touchless acquisition environments,a bimodal palm image alignment approach is proposed based on binocular stereo vision.First,a bimodal imaging system with distance sensors is designed,capable of real-time palm image acquisition.Then,an algorithm based on constraint minimization is proposed to calculate the attitude parameters of ranging sensors.Finally,a bimodal palm alignment method is designed based on the geometric relations between the measured point,the ranging sensor and the binocular cameras.After calibrating system parameters,bimodal palm images can be aligned in real time according to their current distances.The consistency of bimodal palms’ spatial positions is robust against distance change.Therefore,the infrared modal can be utilized to improve the accuracy and stability of visible palm ROI positioning.Compared with existing alignment schemes,this system has the advantages of high precision,fast speed,low computational overhead,and small size.(2)To address the image distortion problem posed by tilted palms under touchless acquisition scenarios,a multi-modal palm image tilt-correction algorithm is proposed based on palm depth information and a multiocular imaging module.Palms are reconstructed in3-D space using multi-point depth information,based on which 3-D ROIs are extracted.After that,standard 2-D ROIs are obtained through perspective transformation.Corrected multi-modal palmprint ROIs correspond to the same palm region,greatly facilitating subsequent palmprint matching.Therefore,the palmprint recognition system becomes more robust to tilted palms and thus can obtain better multi-modal recognition performance.Under the condition of a certain false acceptance rate,the proposed system can achieve a lower false rejection rate compared with acquisition systems without tilt correction alignment,which can effectively improve the user experience of touchless palmprint recognition on the premise of ensuring security.Then,based on the corrected palm image,the recognition mechanism of palmprint is further studied.Through the analysis of palm images in different scenarios,the influence of the system light source intensities on the accuracy of palmprint recognition is explored.Finally,the touchless palmprint recognition system is superior to the existing schemes in the balance of acquisition speed and recognition accuracy.(3)To alleviate the difficulty of locating the ROI from real-world touchless palmprints that affected by palm rotations,translations,scale variations,finger closures,and complex backgrounds,a robust palm keypoint coordinate regression model is proposed.First,a multi-scale Transformer network is used to obtain information of the palm region.Second,the posture of the palm is adaptively adjusted based on the STN model,during which process the palm keypoint region was magnified.Finally,the palm keypoint coordinates are obtained using multi-scale palm information.Additionally,a traditional palmprint localization algorithm is modified to annotate palm regions and edges under monochromatic backgrounds automatically.Training samples with complex backgrounds,different palm postures,and various image qualities are automatically synthesized using image enhancement techniques,improving the generalization performance of the model in different scenarios.Experimental results show that the proposed network structure and training method have high robustness to complex backgrounds and various palm postures while showing strong generalization ability and ROI positioning performance when transfer testing.Compared with exiting methods,it achieved the best loccalization success rate and the lowest localization error in cross-device testing.(4)To addreess the distinctiveness reduction issue of the palmprint image caused by the change of light intensities in touchless acquisition environment,a high-precision palmprint feature encoding and recognition method based on deep convolutional network and multi-feature fusion is designed.First,a Gabor convolutional network with learnable shape parameters is designed by combining the directional Gabor filter and the deep learning optimization framework,which can efficiently explore the order relationship between directional responses,and based on this,a multi-scale palmprint directional feature encoding module is proposed.The encoding module improves the robustness of the model to illumination changes through coarse-grained features.Then,the enhanced finegrained texture features of the palmprint are extracted through the deep convolutional network,which improves the identification ability of palmprint feature encoding and realizes the high-precision recognition target.Finally,based on the directional Gaussian group convolution network,the robust palm vein pattern extraction module is realized,and the recognition accuracy of the model is further improved by the fusion of the coarse and fine-grained features.The experimental results show that the method proposed in this paper has achieved the best performance on average than the existing methods on multiple datasets,and achieved the minimum equal error rate of 0.323% on the newly established palmprint data set with illumination changes.