Image Deblurring Based On Generative Adversarial Network

As one of the important carriers of information,images contain abundant information.However,in the process of image acquisition,processing,transmission,or preservation,there may be the impact of camera shake or other uncontrollable factors,resulting in different degrees of image degradation,such as blurring and noise,for subsequent image processing or the application causes trouble.Therefore,how to reconstruct clear images from blurred images has become a hot research direction in the field of image processing.To this end,this dissertation uses generative adversarial networks to conduct research on image deblurring.The main work includes:(1)Current generative adversarial networks are constructed by convolutional neural networks,but because the convolution operation is a local operation in space,the correlation between long-range features cannot be efficiently captured,resulting in the generation of the effect of images with complex geometric structures not satisfactory.To solve this problem,an image deblurring model based on attention mechanism to generate adversarial networks is proposed.First,the spatial attention mechanism is introduced based on non-local operations.When calculating the output of each position of the feature map,it is no longer only calculated with local neighborhood pixels,but the weighted average of all positions is taken as the response of the output position;Secondly,the softmax operation further highlights the weights of important positions to complete the spatial attention operation;then,considering the correlation between the feature map channel dimensions,the channel attention mechanism is introduced,and the weight of each channel is learned through the global pooling operation.Thereby generating attention in the channel domain;Finally,the spatial attention and channel attention mechanisms are fused to learn the correlation between features,thereby solving the convolution operation to a certain extent due to the size of the convolution kernel that is unable to capturing the correlation between long-range features,can effectively improve the image deblurring effect.(2)Current image deblurring methods based on deep neural networks mostly use residual learning to improve gradient propagation in model design,so that the deblurring performance can be improved by increasing the model size.However,there are some redundant layers or paths in the residual network,which limits the model's representation ability.To solve this problem,a generative adversarial network based on channel weighted fusion is proposed and used to solve the image deblurring problem.First,the weighting coefficients are introduced into the jump connection branch and the nonlinear mapping branch in the residual module,so that the network can learn the importance of each channel by itself,and fully exploit the model's representation ability;second,introduce the spatial attention in chapter 3 mechanism to solve the problem that the convolutional layer is limited by the size of the convolution kernel and cannot capture the long-range correlation;finally,the dimensionality reduction of the output feature map through the 1 ?1 convolutional layer to reduce the amount of calculation and at the same time The role of module input and output fusion.Experimental results on the GOPRO_Large dataset show that the proposed model can effectively improve the image deblurring performance.The dissertation has 21 figures,7 tables and 106 references.