Image Inpainting Based On Generative Adversarial Networks

Digital image inpainting belongs to the research topic of the intersection of computer vision and graphics.The process is that the computer uses the existing area information in the damaged image to repair and fill the missing area of the image according to a certain repair rule.The primary goal of the technology is to repair damaged images so that the observer is unaware of the damage to the image.Nowadays,digital image restoration technology has been applied in many fields such as cultural relics protection,film and television special effects production,virtual reality and old photo restoration.In this paper,an improved algorithm is designed based on pix2 pix for the problems of poor repair effect and large difference between repair image and target image.The innovation of the algorithm is that the joint overall matching and overall true and false network to complete the image restoration.It not only improves the defects of other surveillance-based image inpainting algorithms,but also improves the accuracy of image restoration.Learn from other supervised image inpainting models,combined with the generative adversarial networks.In order to extract more residual information in the damaged image,the generator uses the U-Net structure to complete the main tasks of image inpainting.The discriminator use an overall matching discriminator and an overall true and false discriminator to accelerate network convergence from different directions.It is specifically used to train the network and is regarded as a secondary network.Considering the ultimate goal of image inpainting and the particularity of the generative adversarial networks,the loss function is defined as the weight matching adversarial loss,the true and false against loss and the L1 loss of between damaged image and target image.In order to avoid the network from escaping to the local optimal solution during the weight update is too large in the training process,the cross-entropy loss with smoothing term is used for the true and false against loss.In the end,the result is better.In order to verify the rationality and effectiveness of the image inpainting algorithm in this paper,two image evaluation methods,peak signal to noise patio(PSNR)and structural similarity index(SSIM),are adopted to compare the PSNR and SSIM of different image inpainting algorithms.The final experimental results show that the value of PSNR and SSIM of the image restoration algorithm in this paper are higher than the other three algorithms in the case of the same damaged area.The experimental pictures and the final experimental comparisons verify the rationality and effectiveness of the image restoration algorithm in the paper.