Research On Image Inpainting Based On Deep Generative Model

Image inpainting is the process of predicting the content of damaged areas by using the information of the known part of the damaged image.With the development of digital image processing technology and the popularity of PCs and smart devices,it is more convenient to edit images,how to repair damaged images to their original state has become a hot research direction of image processing.To a certain degree,the image repair algorithm has achieved a lot so far,however,there are still some problems: Firstly,Many image repair methods use the principle of thermal diffusion to spread the edge pixels into the damaged area along the direction of the image iso-line.This method is excellent for the repair of fine scratches and cracks,but when the damaged area becomes wider and larger,this method cannot do anything.Secondly,other methods that can repair large damaged areas have limited perception and expression ability,they cannot effectively understand the contents of damaged areas and fully display them,even if the contents after repairing are consistent with the original image,the generated image is blurred and the edge fusion is abrupt.In view of the problems above,in this paper we propose a depth generation algorithm based on the Deep Convolution Generative Adversarial Networks(DCGAN),combining the residual principle and attention mechanism,to improve the repair efficiency.Focusing on the ability of network perception and expression,this paper modifies the DCGAN generating model network into a convolutional neural network with encoder-decoder structure,to compensate for the information loss caused by the down sampling of the neural network.Skip Connection is introduced to connect the corresponding stages of the encoder and decoder of the generating model network,and transfer the output of part of the encoder layer to the corresponding layers of decoder.Through comparative experiments,the new algorithm not only improves the effect of repairing large area damaged images,but also reduces the difficulty of depth model training to a certain extent.At the same time,the new algorithm has better tolerance for the size and damage of the restored image.Cognitive psychology holds that when people look at an image,their attention is only focused on the part of the image.Therefore,when predicting the value of a point in an image,the influence of different positions of the image is also different.According to this principle,in order to further improve the perception and expression ability of the algorithm,attention mechanism is introduced to weigh different regions and features of the image from spatial and channel dimensions,to enhance the useful features and suppress useless features,to improve the effect of image generation.At the same time,perceptual loss and Total Variation loss are introduced to constrain the generated model,which makes the edge of the repaired image smoother.Finally,experiments on CelebA dataset show that the new algorithm has better performance on face image data set than the traditional algorithm and the basic algorithm without attention mechanism.