Research On Image Super-resolution Method Via Deep Neural Networks And Sparse Regularization Constraint

Image super-resolution is a kind of image processing technology that can improve the visual effect of a low-resolution image.The image super-resolution technology reconstructs a high-resolution image with more details and better visual effects by a mathematical model between a low-resolution image and high-resolution image.With the development of deep learning and machine vision in recent years,image superresolution methods based on neural network have achieved great progress,and attract great attention from researchers.There are a lot of feature redundancy in the neural network,and the network parameters fail to effectively express the mod el.Those questions will seriously affect the performance of the neural n etwork,and lead to bad reconstruction result of high-resolution images.some methods to effectively solve the feature redundancy and the effective expression of parameters are proposed.The main research contibutions of this thesis are listed below:This thesis first analyzes the image super-resolution model,and secondly introduces a typical image super-resolution method framework based on convolutional neural network.Starting with the composition,the network model,structure and network optimization of convolutional neural network are discussed.Then,the sparse regularization methods include unstructured sparse regularization and structured sparse regularization are introduced.Next,several typical regularization methods which applied in neural network is introduced,and their advantages and disadvantages are analyzed.Finally,two typical types of regularization methods have been detailed.Aiming at the problem of a large number of parameters features redundancy between parameters and parameters,filters and filters,channels and channels in convolutional neural network,a method of image super-resolution algorithm based on convolutional neural network structured sparse regularization constraints is proposed.The proposed algorithm can correctly select important parameters and parameter groups,and make full use of the structured regular powerful oracle property of the group bridge.Eliminating the redundant features between parameters and parameter groups can enhance the generalization ability of the network,improve the sparsity of the network,and improve the performance of the network to obtain high-quality reconstructed images.Experiments verify the effectiveness of this method.In view of the redundancy of network parameters and the failure of network parameters to express network model effectively.By exploring the constraint of L0 model,the constraint of L0 model has the optimal sparse effect on the model.However,the constraint of L0 model has the disadvantage of being discrete and non-differentiable.Therefore,an image super resolution algorithm based on the L0 regular constraint of convolutional neural network is proposed to solve this problem.The proposed algorithm constructs a smooth continuous random distribution by used continuous relaxation of discrete random variables.This method can make it continuous and differentiable by approximate L0 norm,and realize L0 optimization of super-resolution network.The proposed method achieves effective expression of network parameters by reducing the redundancy of super-resolution network parameters on the model.Experiments show that,under the premise of ensuring a high degree of sparsity,my proposed method can achieve good results in both subjective and objective evaluation.