| Finger vein recognition is a biometric technology with a high degree of security,speed and accuracy and has a wide range of applications.During image acquisition,the finger vein images can be blurred due to the external environment and hardware devices,making subsequent recognition difficult.In order to ensure that the quality of finger vein images meets the subsequent recognition requirements,this thesis uses deep learning techniques to achieve super-resolution reconstruction of finger vein images and proposes two image super-resolution reconstruction methods for different target requirements.The main research work and contributions of this thesis are as follows:(1)To address the problem that the same network structure is used to extract high and low frequency information in image super-resolution reconstruction method based on convolutional neural network,which leads to a large amount of network computation,this thesis proposes a super-resolution reconstruction method for finger vein images based on frequency division and deep separation of residual attention.This method constructs a model of high and low frequency separation,using networks of different complexity to extract high and low frequency features of the finger vein image respectively,and also use depthwise separable convolution instead of ordinary convolution to speeds up the extraction of image features and reduces calculated volume.In the high-frequency network,deep separation residual attention network is built to enhance the quality of reconstructed finger vein images.In this network,coordinate attention mechanism is used to enhance the extraction of key information in the feature map,and use jumping connections to integrate the features of each layer.In the low-frequency network,simple residual blocks are used for feature extraction to reduce the calculated volume.The experimental results show that the reconstructed finger vein images have better visual effects and higher objective evaluation index values,and the network has a smaller calculated volume,which is useful in the task of super-resolution reconstruction of finger vein images.(2)To address the problems that reconstructed image generate false texture information and unstable network training in the super-resolution reconstruction method based on generative adversarial networks,this thesis proposes a method based on a double discriminator generative adversarial network for finger vein images super-resolution reconstruction.Firstly,This method uses the lightweight network proposed in Chapter 3 in the generator to reduce the network complexity in general.Secondly,a dual discriminator is used to discriminate the generated finger vein images globally and locally respectively,so that the reconstructed image’s texture details can be enhanced by the joint action of the two discriminators.Besides,the SN layer is used to constrain the performance of the discriminator,and uses a relative average discriminator in the local discriminator to evaluate the relative authenticity of the generated image,so as to enhance the stability of the network training.Finally,the texture loss term is introduced into the loss function to enhance the constraint effect on the texture-deficient finger vein images,so that the vein texture detail can be further recovered.The experimental results show that the local texture details of the reconstructed finger vein images are clear and the network training is more stable.(3)Based on the image super-resolution method proposed in Chapters 3 and 4,the finger vein image super-resolution reconstruction system is designed and implemented,and the functions of user management,image reconstruction and image management are realized.In order to verify the effectiveness of the system,finger vein images were selected to test the system.The system test proves that the system can meet the demand of high-quality finger vein image reconstruction.This thesis includes 35 figures,12 tables and 80 references... |
PDF Full Text Request