Perception-Aware Contrastive Learning-Based Progressive Image Inpainting

Deep learning-based image inpainting has demonstrated impressive performance in improving visual quality of the inpainted content. However, existing methods still generate unpleasant content, especially in cases with large broken areas. To solve this problem, we propose a perception-aware contrastive learning-based progressive image inpainting approach, which is implemented in an iterative way to refine the details of the inpainted images. In order to produce results with improved visual quality, a contrastive learning strategy coupled with a perceptual loss is adopted in the refinement process for model optimization. The experimental results demonstrate the excellent performance of the proposed method, as proved by up to 5.48% gain on the Frechet inception distance (FID) and 1.87% gain on the learned perceptual image patch similarity (LPIPS), compared with the state-of-the-art approach.