Image Blind Motion Deblur Algorithm Based On Generative Adversarial Networks

In the digital multimedia era,with the increasing popularity of imaging equipment such as mobile phones,cameras.image has become an more and more important carrier.But,in the imaging process,it is difficult to maintain a relatively static state between the image device and the objects,the image motion blur is caused,and these blurred images have caused a lot of trouble to the information transmission.However,in the fields of daily life,traffic safety,medicine,military reconnaissance,etc.,it is particularly important to obtain a clear image.Therefore,this research topic has important practical significance for many fields.This dissertation first introduces the progress of image deblurring and related theoretical knowledge.Then,under the condition that the traditional image deblurring method has shortcoming in estimating the clear image,the in-depth study of the Generative Adversarial Networks(GAN)and Convolutional Neural Networks(CNN).An image motion blind deblurring algorithm based on GAN and an image motion blind deblurring algorithm based on cyclic multi-scale GAN are proposed.The main work of this dissertation are listed as follows:1.An image blind motion deblurring algorithm based on GAN is proposed.Considering the gradient dispersion problem in the training process of deep networks,we introduce the residual to train the generator.Loss of perception and loss of structural similarity are introduced as constraints in the loss function to improve the results.The algorithm was evaluated on the dataset Gopro,a complex large-scale motion blurred image set,and a good deblurring effect was obtained with an average PSNR of 29.3 and an average SSIM of 0.946.2.Considering that the teaditional method has different blur kernel size,it will affect the restoration effect.Referring to the multi-scale method of estimating the blur kernel step by step in the restoration process,An image blind motion deblur algorithm based on cyclic multi-scale GAN is proposed.This algorithm reduces the influence of blur image blur kernel size on the restoration effect,and has a better restoration effect on the blurred image with large size of blur kernel.In order to balance the image features and the complexity of the network structure,the cyclic multi-scale codec network is used as a generator to share the network weights at different scales of the image,which significantly reduces the number of network parameters,thereby reducing the complexity of training and improving stability.To overcome the information loss caused by the deepening of the network depth,the residual learning module is introduced to improve the codec network.considering images with significant edge information,we used the 2l norm constraint on image gradients.The results show that the algorithm can achieve better deblurring effect than the state-of-the-art algorithms.,and the average PSNR is 30.28 and the average SSIM is 0.945 respectively,blurred images have better recovery effects.