Research Of Image Inpainting Algorithm Based On GAN

In recent years,with the rapid development of artificial intelligence and computer vision,image inpainting has become an important research field,which is widely used in photography,security,medicine and other industries.The effect of traditional image inpainting algorithm in practical application is not ideal.In recent years,deep learning has achieved remarkable results in the field of image processing.With its powerful feature expression and learning ability,Generative adversarial networks gradually replace the traditional image inpainting methods based on pixel diffusion and patch class.This thesis analyzes the research status of image inpainting technology at home and abroad,studies the structure information(edge contour)and color information(color content)of the image,and completes the image inpainting task through the improved generative adversarial networks.The main work of this thesis is as follows(1)Research and improvement of edge information inpainting model for incomplete regionComplete and accurate edge information can give rich and key feature information to content inpainting network,and make image inpainting result more consistent with context semantic information.This thesis proposes a multi-scale content attention mechanism,which uses the features of known patches as convolution filter to process the pre generated patches,and uses multi-scale patches to achieve the consistency of basic features and overall style between the incomplete generation area and the surrounding background area.(2)Research and improvement of content information inpainting model for incomplete areaOn the basis of edge features,the generative adversarial networks of content generate RGB content image to get a repair image.In this thesis,Ada IN residual block is proposed,which not only solves the problem of network gradient vanishing,but also changes the data distribution on the feature map level to realize the realistic generation of image details.In addition,this thesis also proposes a multi region discriminator,which satisfies the completion of any missing shape,and pays attention to the style and semantic consistency between the inpainting result and the complete image.(3)Design and implementation of image inpainting model based on generative adversarial networks.On the basis of the above research,this thesis improves Nazeri's work,designs and implements an image inpainting network model based on generative adversarial networks,which is mainly divided into two stages: 1)edge generation;2)content generation.The core idea of the inpainting model is to guide the generation of color content by the edge contour,so that the inpainting results have reasonable structural changes in the global,and have fine details in the local,to meet the global and local consistency.In this thesis,we compare the proposed network model on Celeba and Places2 image datasets,and verify the effectiveness of edge information,Ada IN residual block,multi-scale content attention mechanism and multi region discriminator.In addition,through the comparative experiments between the inpaining model of our thesis and the related work,it is proved that the image inpainting algorithm in this thesis can generate more realistic and more visually consistent images.