Image Restoration Based On Generative Network Model

Image restoration aims to restore a clean image from the degraded observation as much as possible.Image restoration has existed for a long time,it is a basic problem in the research of computer vision.However,the mathematical expression of image restoration and the universality of the algorithms have not been resolved.These problems hinder the further research and development of image restoration.Recently,the in-depth researches of deep learning provide a solution to image restoration.It can be proved that more powerful and comprehensive prior information and more universal unsupervised algorithms will significantly improve the algorithm's resilience.The research in this paper is based on the theoretical of image restoration,which uses generative network and unsupervised algorithms.This paper is focus on image inpainting and image demosaicing tasks:(1)Many researchers have successfully used generative networks in image restoration.By analyzing their successful experience,we propose a reversible generative network in the wavelet transform domain and the pixel domain.In the wavelet transform domain,we introduce the discrete wavelet transform,which changes the input image from the image domain to the time-frequency domain.The final time-frequency domain image is one-fourth of the original size,and the number of channels is four times of the original.At the same time,the scale transformation of the pixel domain transforms the input image into a quarter of the original image through down-sampling,the number of channels is expanded to four times of the original image,and then enters the network as a new input.Finally,we enter the two transform domains into different reversible generating networks and make their respective outputs obey Normal distributions.(2)The biggest difficulty of image demosaicing is that the image loses too much information.In one research mode,the image loses two-thirds information of the original image,so the solution is difficult.In this research,we propose an unsupervised learning algorithm that uses a progressive solution algorithm that combines image domain and gradient domain.Compared with other algorithms,our algorithm pretrains the generation network to obtain the prior information of the image,and uses the prior information in the final image solution process.In summary,the two image restoration tasks studied in this paper both use generative networks as tools and unsupervised learning algorithms as the core,and use the prior information of generative network learning for the final image restoration.