Research On Finger Vein Image Capturing And Enhancement Based On Deep Reinforcement Learning

In recent years,due to the impact of the global pandemic,higher demands are placed on public health,and due to the advantages of non-contact collection and unique and difficult-to-fake characteristics,vein recognition has been regarded as a research hotspot in the field of biometrics.However,due to issues such as blurring,noise,and missing vein patterns in the actually collected vein images,the subsequent matching and recognition are seriously affected,and images with multiple problems cannot be effectively solved by existing enhancement methods.In this paper,the collection of vein images,deblurring or damage restoration of images with single issues,and adaptive enhancement of images with multiple problems have been used to propose a series of innovative works.Specifically:(1)A new method for vein image collection is proposed.Jetson Nano is used as the main control module to improve computing power,and PCA9685 is designed as a module to achieve precise control of the infrared light source.A new structural case is designed to avoid uneven illumination.Additionally,a brightness evaluation and noise point evaluation method is proposed to preliminarily determine the intensity of the infrared light source;(2)A novel method for vein image deblurring,which improves Deblur GAN-V2,is proposed.In this method,Inception-Res Net-V2 and FPN modules are combined in the network to improve the vein feature extraction ability and deblurring performance.Improved losses composed of global and local image discriminators,vein minimum binary loss,vein feature map perceptual loss,and adversarial loss are jointly used to constrain and train the network,guiding the enhancement process of the blurred finger vein images;(3)A novel two-stage image damage restoration method based on guided finger vein texture features is proposed.In the first stage of this method,the finger vein texture features are used to provide constraints and guidance for the restoration of the original image in the second stage.Style transfer techniques are introduced to fully utilize the vein features to transfer the texture into a clear finger vein image.After restoration,the missing finger vein texture is more coherent and reliable in the repaired areas of the image;(4)A novel finger vein image adaptive enhancement method based on improved DQN,called DRL-FVRestore,is proposed.In this method,the enhancement task that can handle various image problems such as image deblurring enhancement,image damage restoration,feature enhancement,and denoising,is considered as the decision-making behavior of the agent.The SE-Res Ne Xt is used to improve the feature network of DQN to enhance the feature extraction capability of finger vein images.The LSTM is used to replace the historical experience pool of DQN,enabling the agent to select enhancement behaviors for image processing continuously and adaptively according to historical information when dealing with various image problems.Combining the above research,a finger vein collection and enhancement system(FVCE)is designed,which realizes the collection,infrared light source control,quality evaluation,and adaptive enhancement operations of finger vein images.Four public finger vein datasets were used to verify the performance of the collection method and the rationality and reliability of the enhancement method.Template matching,feature point matching,subjective and objective quality evaluation of finger vein images were regarded as evaluation indicators.The experiments show that the proposed collection method can rapidly and efficiently capture finger vein images with clear patterns.Regarding the single-image problem,the proposed deblurring and defect repair methods show an average reduction of 4.31% and 1.71%,respectively,in the EER value compared to other methods.For the multi-image problem,the proposed DRL-FVRestore method shows an average reduction of 3.98% in the EER value.