Research On Image Deblurring Method Based On Generative Adversarial Networks

With the highly developed information technology,smartphones and digital cameras have been widely available.As the major information carrier in today's social networks,image contains a wealth of information.It can record facts,share people's lives and convey their emotions,being widely used in daily life.As a method of image restoration,the study of motion deblurring is becoming more potential for industrial application.In the field of image deblurring,data acquisition is a tough problem since we cannot capture image pairs of corresponding sharp and blurred images.Thus,most scholars have focused on studying methods for generating synthetic blurred images and training deblurring algorithms on them.However,the distribution characteristics of synthetic datasets are different from those of realistic blurry images captured from real world,leading weak generalization ability in the algorithms trained on them.When these algorithms are applied to realistic blurry images,their effectiveness would decline sharply.In order to promote the practicability and the robustness of deblurring algorithm,this thesis propose a novel blind image deblurring method based on generative adversarial networks.The method has the following advantages:1.The proposed algorithm can be trained on unpaired image datasets,thereby realistic blurry images can be used in the process of learning.2.The proposed algorithm follows the idea of image-to-image translation,redefining the image deblurring problem as the transformation from blurry domain to sharp domain.3.The proposed algorithm is consisted of two generative adversarial networks.Deep residual networks,Wasserstein GAN,cycle consistency,perceptual losses and other cutting-edge deep learning technologies are used to optimize the model,which can reconstruct blurry images to sharp,nature and realistic images.The proposed method achieves a strong performance on several datasets and evidently outperforms CycleGAN,making it promising to solve the motion blur in the wild.