Parallel adaptive guidance network for image inpainting

Abstract Motivated by human behavior, dividing inpainting tasks into structure reconstruction and texture generation helps to simplify restoration process avoid distorted structures blurry textures. However, most of are ineffective for dealing with large continuous holes. In this paper, we devise a parallel adaptive guidance network(PAGN), which repairs enriches textures through branches, several intermediate-level representations in different branches guide each other via the vertical skip connection filter, ensuring that branch only leverages desirable features another outputs high-quality contents. Considering larger missing regions are, less information is available. We promote joint-contextual attention mechanism(Joint-CAM), explores between unknown known patches measuring their similarity at same scale scales, utilize existing messages fully. Since strong feature representation essential generating visually realistic semantically reasonable contents regions, further design attention-based multiscale perceptual res2blcok(AMPR) bottleneck extracts various sizes granular levels obtains relatively precise object locations. Experiments on public datasets CelebA-HQ, Places2, Paris show our proposed model superior state-of-the-art models, especially filling