Resumen
Image inpainting aims to synthesize missing regions in images that are coherent with the existing visual content. Generative adversarial networks have made significant strides in the development of image inpainting. However, existing approaches heavily rely on the surrounding pixels while ignoring that the boundaries might be uninformative or noisy, leading to blurred images. As complementary, global visual features from the remote image contexts depict the overall structure and texture of the vanilla images, contributing to generating pixels that blend seamlessly with the existing visual elements. In this paper, we propose a novel model, PA-DeepFill, to repair high-resolution images. The generator network follows a novel progressive learning paradigm, starting with low-resolution images and gradually improving the resolutions by stacking more layers. A novel attention-based module, the gathered attention block, is further integrated into the generator to learn the importance of different distant visual components adaptively. In addition, we have designed a local discriminator that is more suitable for image inpainting tasks, multi-task guided mask-level local discriminator based PatchGAN, which can guide the model to distinguish between regions from the original image and regions completed by the model at a finer granularity. This local discriminator can capture more detailed local information, thereby enhancing the model?s discriminative ability and resulting in more realistic and natural inpainted images. Our proposal is extensively evaluated over popular datasets, and the experimental results demonstrate the superiority of our proposal.