Generative Memory-Guided Semantic Reasoning Model for Image Inpainting

The critical challenge of single image inpainting stems from accurate semantic inference via limited information while maintaining quality. Typical methods for train an encoder-decoder network by learning a one-to-one mapping the corrupted to inpainted version. While such perform well on images with small regions, it is challenging these deal large area due two potential limitations. 1) Such paradigm tends overfit each training pair images; 2) inter-image prior knowledge about general distribution patterns visual semantics, which can be transferred across sharing similar not explicitly exploited. In this paper, we propose Generative Memory-guided Semantic Reasoning Model (GM-SRM), infers content regions based only known image, but also learned reasoning priors characterizing generalizable between images. particular, proposed GM-SRM first pre-learns generative memory whole data learn different patterns. Then are leveraged retrieve matching semantics current during inpainting. used guaranteeing pixel-level consistency, our favorable performing high-level reasoning, particularly effective inferring area. Extensive experiments Paris Street View, CelebA-HQ, and Places2 benchmarks demonstrate that outperforms state-of-the-art in terms both quality quantitative metrics.