Research On Image Motion Blur Blind Removal Algorithm Based On Deep Learning

Image is an important carrier for people to obtain information,so digital image processing technology is booming.In the past few years,the emergence of generative confrontation network provides a new direction for many scholars at home and abroad in image deblurring research.Many excellent deblurring models have been proposed and achieved good results,but there are also some problems,such as chessboard artifacts in the generated image,lack of details in the generated image,poor generalization ability of the model and so on.Aiming at these unsolved problems,this paper studies the blind removal algorithm of image motion blur based on deep learning.The causes of bad image quality and chessboard artifacts in the generated image are analyzed,and the network structure and loss function are improved and optimized:the network generator structure is improved,and the deconvolution module and pixel recombination module are used as the up sampling structure in the generator network;the network loss function is optimized to combat the loss plus pixel domain The form of content loss plus feature domain content loss is taken as the total loss function of the network.The loss function of WGAN-GP is used to resist the loss.The content loss in pixel domain is L1 loss,and the content loss in feature domain is in the form of perceptual loss.The perceptual loss is calculated by extracting image features from VGG network and Dense NET network.Therefore,a blind image motion blur removal method based on generative countermeasure network is proposed.Aiming at the problems of poor generalization ability of network model based on paired data set training and difficulty in practical scene application,a blind image motion blur removal method based on cyclic generation anti network is proposed.By using the method of image translation,the problem of image deblurring is transformed into the problem of mutual conversion between the fuzzy domain and the clear domain,and the network structure and loss function are improved and optimized,so that it can be trained in both paired and non logarithmic data sets.Finally,the two methods are verified by experiments.The experimental results show that the two methods have better ability of removing motion blur.The objective evaluation index of this method is improved,and the chessboard artifacts in the generated image are eliminated more effectively,which has a good application prospect.