Research On Image Inpainting Based On Generative Adversarial Networks

Image inpainting technology refers to filling the missing part of the image,which is widely used in military,industrial,archaeological and other fields,and thus becomes an important research direction in the field of computer vision.The traditional image repair method is mainly based on the method of pixel value filling,which refers to searching for a part similar to the edge area to be repaired in the intact area of the image as a patch filling.But when the image has complex semantic information or the missing area is slightly larger,the inpainting effect of traditional methods is poor.After the concept of generative confrontation network and deep learning was proposed,the task of regular mask image repair in large areas has made great progress,but in most application scenarios,image missing areas are often random and irregular.The inpainting results often have problems such as artifacts,chromatic aberration and unrealistic details.In order to solve the above problems,this paper proposes a coarse-to-fine twostage generative adversarial network image inpainting model for the repair task of irregular masks,which improves the quality of image repair to a certain extent.The main work and contributions of this paper are as follows:(1)A two-stage generative adversarial network image inpainting model based on multi-scale feature fusion is proposed.In the first stage,the gated convolution stacked network is used to construct the network structure.In the second stage,multi-scale feature fusion and confrontation training are used to clarify the blurred image,optimize details,and improve the repair quality.Using spectral normalization techniques in the discriminator to stabilize the training of generative adversarial networks.Experimental results show that when repairing face images with irregular masks,mask traces can be effectively removed,and content that is consistent with the global image structure and semantically coherent is generated in the missing area,and a good repair effect is obtained.(2)Propose a dual discriminative generative adversarial network image inpainting model based on gated convolution and SENet.The inpainting process consists of coarse restoration and refine restoration.Firstly,input the damaged image mask into the coarse network stacked by several gated convolutions,add channel attention during upsampling,and combine L1 reconstruction loss to obtain a preliminary repair map;then,input the preliminary repair map into the fine network,the fine network is composed of several gated convolution blocks and channel attention blocks,combined with reconstruction loss,perceptual loss,and adversarial loss to improve important features and details,and cover the intact area of the damaged image on the repair map of the fine network as the completed result;Finally,the double discriminant network structure is used to improve the inpainting quality.The proposed model was tested on the celeb A data set,and the inpainting result reached 27.39 in peak signal-to-noise ratio(PSNR),which was 6.74% higher than partial convolution,and reached 0.9216 in structural similarity(SSIM),compared with partial convolution,it has improved by2.95%.Experimental results show that Squeeze-and-Excitation attention mechanism and dual discriminative structure can improve the details of image inpainting.