Research On Image Restoration Based On Generative Adversarial Networks

Image blurring problem often leads to the decrease of the accuracy in postprocessing.Therefore,the restoration of blurred images has been a hot topic in the field of computer vision.Image restoration can be categorized as a complex data generation tasks,while the Generative Adversarial Networks is good at solving such tasks.However,there are still some problems in the current research: firstly,the generated image will be missing some details.Secondly,the model degradation phenomenon often occurs.Thirdly,the paired images are difficult to collect,resulting in the network not being robust.In view of the above problems,this thesis focuses on designing an image restoration network for unpaired image sets.The main research work of this thesis is as follows:(1)In order to solve the problem that the generated image loss of partial detail information,the residual network is added to the generator model,thus the network only learns the difference between the blurred and clear image,which reduces the image detail loss rate and generates a more detailed image.The experimental results show that the improved method can generate higher quality images and effectively improve the restoration ability of the network.(2)In order to solve the problem of model degeneration in the process of network training,Wasserstein distance is introduced as a measure of the difference between the generated image distribution and the real image distribution.Then add an improved cascading model in the network to generate multi-scale features and enhance the learning ability to image features.The experimental results show that the method in this thesis effectively solves the problem of training collapse,and makes the model generate high-quality images while ensuring the stability of network training.(3)In order to solve the problem that the paired images are difficult to collect,the Dual Learning is introduced into the image restoration method based on the Generative Adversarial Networks,this thesis proposes an unsupervised image restoration based on Dual Adversarial Learning.The experimental results show that the network training on the unpaired image data set can still maintain a good restoration effect.It shows that the method of this thesis solves the dependence of traditional methods on paired data sets.