Image Inpainting Generation Adversative Network Based on Attention Mechanism

Abstract

Image inpainting is an essential issue in image processing, which aims to use computer vision technology to restore damaged or lost parts in degraded images automatically. It is widely used in film and television special effects production, such as image editing, digital heritage protection, and other fields. To aim at the problem of image inpainting with relatively complex image structure and high requirement for local image region repair, this paper proposes a deep convolution generation adversative network image repair method combining the local and global refinement network (LG-net) and Convolutional Block Attention Module (CBAM). Firstly, the idea of the generative adversarial network is adopted to capture semantic information of the missing part of the image by using the encoder-decoder network of skip connections. Secondly, to enhance the local repair effect, we built a local development network based on a small receiving field, which can also weaken the effect of distant undesired completion results. Thirdly, we use Convolutional Block Attention Module to obtain the important degree of features, enhance the ability to guide feature information, and further refine the global feature. Finally, the stability of the network is enhanced by introducing the l <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> loss, weighted reconstruction loss, perceptual loss, style loss, and TV loss. For the CelebA-HQ dataset, the proposed method obtained a PSNR value of 30.98 and an SSIM value of 0.983, respectively. Moreover, for the ParisView dataset, the proposed method achieved PSNR values of 29.60 and SSIM values of 0.986. Experimental results show that the proposed method is better than the current mainstream image restoration techniques in quality and quantity.