Old-Photo Restoration with Detail- and Structure-Enhanced Cascaded Learning

Restoring old photos with unknown and complex degradations remains proverbially meaningful but challenging today. The related image inpainting and image restoration methods can not address structured and unstructured defects in old photos simultaneously, which will produce over-smoothed, lacking well-preserved structure and outdated results. In this paper, we propose a cascaded two-stage old photo restoration method: the coarse stage mainly restores unstructured defects by learning the mapping function between the degraded image domain and the clean image domain; the refinement stage focuses on filling in missing or damaged regions with a multi-scale residual dense edge restoration network and a Fourier inpainting network. Particularly, we embed simple yet effective feature extraction modules to retain detail features in the coarse stage and extract the structure information from the edge map in the refinement stage. Experiments show that, with dedicated designs, our approach surpasses the-state-of-art old photo restoration methods qualitatively and quantitatively.