Research On Image Inpainting Based On Improved Generative Adversarial Network

As a branch of computer vision-image restoration,it has always been an important research field pursued by researchers,especially with the development of deep learning,the research enthusiasm for image restoration continues to heat up,reaching an unprecedented height.Today,many well-known image restoration methods are proposed based on Generative Adversarial Network(GAN).Nevertheless,there are more or less diffculties for the existing image restoration methods.Based on this,this thesis proposes two image inpainting models based on improved generative adversarial networks.The research contents are as follows:First,in the process of image inpainting,since the texture and color of the image synthesized by the generator may be inconsistent with the real image,this is frequently called false texture.In order to reduce the damage of unreasonable pixels to the generated image,this thesis proposes a network model(FAM-GAN)based on Fake-texture Attention Map(FAM).This mechanism is dedicated to training the fake texture area and generating a fake texture map,comparing it with the gray pixel difference map generated between the real image and the repaired image to get the error,and feeding it back to the network to eliminate the error of the feature map in the generator Pixel inconsistency,generate a more realistic picture texture,so as to achieve the purpose of image restoration.Second,in the field of image restoration,the proposed methods basically have a large receptive field,which is suitable for repairing images with relatively concentrated missing areas,but provided that the missing areas are scattered,the repair effect of the image is not good,that is the results of a larger receptive field may not be ideal.For this reason,from the perspective of receptive field and mask update,this thesis proposes a three-stage image inpainting model based on global and local refinement,which replaces traditional convolution with gated convolution.The specific process is as follows: first,input the original image and mask data set into the rough reconstruction network(small U-Net network)to obtain a rough image repair result;secondly,enter the local refinement network(network model with a small receptive field),perform local refinement on the missing parts of the image;finally,input the above results into the global refinement network(a network model with a great receptive field),perform global refinement on it,and generate the final result.By comparing the experimental results of three classic image inpainting approachs on the Paris Street View Google Street View dataset and the Celeb A-HQ dataset,The proposed method performs well on two popular publicly available inpainting datasets.