Research On Image Inpainting Based On Deep Neural Network

Art restoration,the technique that aims to revert deterioration of artefact to close to the human visually plausible manner,has occurred in hundreds years ago.Digital image in-painting,a relatively young technique to art restoration,has drawn attention from many people.Image in-painting technique is used to fix the missing information of image through the neighbor pixels of missing information.Image in-painting could apply in many areas,for examples digital image object removal,image reconstruction,text removal,video restoration,movies disocclusion and so on.Several approaches have been proposed relative researchers to solve the problem.There are five types of method based on different solutions,which based on partial differential equations,textures,samples,mixture and deep learning,to solve the image inpainting problem,and differently perform.Firstly,this article introduces deep convolution network and generative adversarial network based on deep learning theory.This article introduces some features of deep convolution network,a classical deep convolution network for image classification task,definition and confirmation of generative adversarial network,and some methods to optimize deep learning models.Secondly,we present a method based on deep convolution network and generative adversarial network to image inpainting.This method uses two different stages to fix missing information of image,first stage for coarsely inpainting and second stage for finely inpainting.Furthermore,two different methods are implemented to finely inpainting in second stage.Finally,the method introduced in this article was compared to other methods in common dataset CelebA,and the experience shown that the method presented in this paper performed better than traditional method and method based on deep learning.