Research On Finger Vein Recognition And Template Protection

In recent years,the global epidemic has led to further developments in biometric technology.As the second generation of biometric technology,finger vein recognition has become a research hotspot due to its advantages such as high security,and living body recognition.Finger vein recognition is an identification technology based on biophysiological features.When a finger is illuminated by near infrared light,the haemoglobin in the vein blood absorbs more near infrared light than the neighbouring tissue,resulting in shadows during the imaging process,which are represented in the image as vein patterns.By analyzing the vein pattern in the image,which is used as a matching reference for authentication.Although finger vein recognition-related research has made some progress,there are still some problems that need to be refined.For example,the variety of collection devices and changing collection environments can lead to complex and changing background noise distributions in the images.Also,finger vein images have a large range of grey-scale gradient pixels in the joint cavity,which makes the traditional gradient operator less robust.Moreover,finger vein images are generally small in data size,which will not provide sufficient training samples,resulting in lower performance of deep learning based recognition systems.In addition,once a biometric is leaked,the security of the recognition system is drastically reduced and this biometric will no longer be available,limiting the biometrics available to the subject.This thesis is an in-depth research on finger vein recognition: image collection,preprocessing,recognition and template protection.The main work and contributions include:(1)Unconstrained vein collection devices,due to the variety of structures and the variability of the collection environment,the captured images all suffer from a large amount of unstable and complex background noise,which affects the system recognition performance.To overcome these problems,the following exploratory work was implemented.Firstly,the acquisition device have homogeneous infrared light source module,which uses triangular concentrated physical structure to concentrate the emitted light intensity in some small regions.Secondly,a V-Vibe finger vein image Region of In-terest(ROI)extraction method was proposed,which converts the removal of complex backgrounds and the difficulty of judging weak and pseudo-edges into a similarity judgement with each pixel point in the background image through background differencing.A new finger vein dataset,Finger Vein-Background Subtraction(FV-BS),was created based on the designed acquisition device.(2)An ROI extraction method based on the finger vein imaging characteristics was proposed.Firstly,considering the finger contour characteristics during imaging,a finger edge search rule was designed and used to obtain the ROI horizontal segmentation reference line;considering the large range of graduated gradient pixels generated by the finger joint cavity during imaging,a joint cavity localization method based on a large receptive field gradient detection operator was proposed and used to obtain a stable and accurate vertical direction segmentation reference line.Finally,the final ROI region is obtained using the segmented reference lines in both horizontal and vertical directions.Different from the proposed V-Vibe method,this method further divides the finger region in the vertical direction,eliminating the influence of finger position or shape on recognition.(3)Insufficient training samples of finger vein images can lead to poor feature learning and weak model generalization,limiting the performance of finger vein recognition systems.In the data-driven aspect,a simple and effective finger vein data expansion strategy was proposed: FV-Mix,which uses fine finger vein image ROIs for grey-scale normalization and linear fusion.On the knowledge-driven side,a Residual Gabor Convolutional Network(RGCN)was proposed to guide the deep learning model to learn the pattern features in the finger vein images,which consists of a residual Gabor Convolutional Layer(RGCL)and a Dense Semantic Analysis Module(DSAM).(4)A cancelable finger vein biometric recognition system was proposed.Firstly,the Block Warping Remapping(BWR)template protection method is designed,which can generate protected feature templates in combination with user externally defined keys.Secondly,the protected templates are corrupted by the finger vein pattern,detail points and other features due to the warping and remapping operations,and the finger vein recognition system is designed to improve the recognition performance by building a Convolutional Neural Network(CNN).The use of BWR-protected feature templates can effectively prevent attackers from acquiring the original features through the recognition model.At the same time,the links between pixels at the distorted block boundaries are diluted,and the unlinkability and irreversibility of template was further enhanced.