Research on Image Inpainting Algorithm Based on Self-Attention Mechanism

Abstract This paper presents a two-path image coloring network based on generative adversarial networks to address structure loss and texture blurring in image coloring. The network uses a modified residual network and adds long-term and short-term attention mechanisms in the decoding stage. The input picture is identified using a patch discriminator. The algorithm is validated using the Places2 dataset and its restoration results are evaluated using image evaluation methods. Experiments show that the larger the occluded region, the better the results. The algorithm improves psnr and ssim values by about 2% and 15% compared to traditional algorithms, and the images repaired are clearer and more realistic.