Image Inpainting Based On Multi-Scale Feature Representation Model

Image inpainting refers to the technology of inpainting or repairing degraded images or damaged pictures.It aims to automatically restore the missing content in the image based on the known content.With the rapid development of network bandwidth technology and intelligent applications,the frequency of using video,image and other streaming media data in the life of modern human society has become more and more frequent,and people’s demand for image processing and image inpainting software tools is increasing.Recently,because of the rapid development of deep learning technology in the field of image processing,the learning-based image inpainting technology has received extensive attention from academia and industry.This paper summarizes three common problems: the restored image edge is inconsistent with the real visual world,the model discards detailed texture information during forwarding propagation,and the model is too large.In order to solve these three problems,this article explored and practiced three different inpainting techniques.The main contributions of this thesis are as follows:(1)The sketch map of the picture guides the inpainting process and makes the restored image more edge information.In this proposed method,the inference process of image inpainting is divided into two stages: sketch map inpainting stage and RGB image inpainting stage.In the second inpainting stage,the repaired sketch map results by the first stage will be inputted to the second deep generative model as a clue.Experiments demonstrate that this technology greatly improves the integrity of the edge information of the restored image.(2)Based on the traditional generative adversarial network,the generator in our second proposal is constructed as a multi-scale feature representation model.The designed network can produce feature representation in two different scales,and it will fuse and restore the information of these two kinds of features in the forward propagation of the network.The inpainting performance of the proposed network is higher than the state-of-the-art inpainting approaches.(3)The feature pyramid network inpainting model is applied to the generator of image inpainting in our third proposal.This network also has a multi-scale feature representation,but its parameters are very large.Therefore,we optimized the feature pyramid network at the macro and micro levels.At the micro level,we propose a unique lightweight residual convolution module.At the macro level,we optimized the overall structure of the network,especially the top-down part of the feature pyramid network model.The entire optimized network can complete the image inpainting task efficiently and quickly under the premise that it can express features of different scales.