Research On Motion Image Deblurring Based On Attention And Adversarial Network

Motion image deblurring technology is an important research direction in the field of image restoration,its purpose is to reconstruct and estimate the clear images of the blurred images caused by camera jitter and object motion during photographing through image restoration technology,which is widely used in safety monitoring,aerospace,and many other fields.In recent years,with the advent of deep learning wave,it has been applied to Deblur motion images and achieved satisfactory results.However,in the current motion image deblurring algorithm based on depth learning,there are still some problems such as insufficient high-frequency detail reconstruction and inability to effectively remove artifacts.In order to further restore the high frequency texture details in the blurred image and improve the quality of the deblurred image,this paper introduces the attention mechanism into the generative adversarial network to extract and enhance the high frequency features,A motion image deblurring method based on attention and generative adversarial network is proposed.The main research work is as follows:(1)Aiming at the problems that the current motion image deblurring network ignores the non-uniformity of the moving fuzzy image and cannot effectively restore the highfrequency details of the image and remove artifacts,A motion image deblurring method based on adaptive residual is proposed based on generative adversarial nets.This method constructs an adaptive residual module composed of the deformation convolution module and channel attention module in the generating network.Among them,the deformation convolution module learns the shape variables of motion fuzzy image characteristics,which can dynamically adjust the shape and size of the convolution kernel according to the deformation information of the image,and improve the ability of the network to adapt to the deformation of fuzzy images.The channel attention module adjusts the extracted deformation features to obtain more high-frequency features of images and enhance the texture details of the restored images.The experimental results show that compared with the existing algorithms,the method can effectively remove blur and reconstruct high-quality images with rich texture details.(2)To better extract global context information,enhance feature expression,and effectively reconstruct the texture details of fuzzy images.A motion image deblurring method based on recurrent cross-attention is proposed.This method is based on the adversarial network and uses a recursive cross-attention module to realize the extraction of global context information in the generated network,to promote the reconstruction of highfrequency texture details of the image.At the same time,to make the generated network better fit the data distribution of clear images during training,the relative discrimination principle is added based on the least-squares confrontation mechanism during confrontation training.Among them,the relative discrimination principle makes the data distribution of clear images as a priori knowledge to restrict the confrontation network by adding the evaluation input data and the data samples of its opposition types more like the true probability,to enhance the fitting ability of the generated network to clear images.Experimental results show that compared with the existing motion image deblurring algorithm based on a deep convolution neural network,the proposed algorithm can reconstruct high-quality deblurring images and improve the numerical index.