Reviving Old Photographs

Abstract

Our Project is titled “Reviving Old Photographs” is focusing on restoring old photos back to life, Through a deep learning method, we suggest restoring antique photos that have suffered serious damage. The degradation in genuine photos is complex, and the difference in domain between artificial images and actual old photos prevents the network from generalizing, in contrast to conventional restoration tasks that can be solved through directed learning. Lever-aging genuine photographs together with a large number of synthetic image pairs allows us to propose a novel triplet domain translation network. We train two variational autoencoders (VAEs) to convert clean photographs into two latent spaces and to change aged photos into two latent spaces, respectively. Additionally, synthetic paired data is used to learn the translation between these two latent areas. Since the do-main gap in the compact latent space is closed, this translation generalizes effectively to actual photographs. In addition, we create a global branch with a partial nonlocal block targeting the structural defects, like scratches and dust spots, and a local branch targeting the unstructured defects, like sounds and blurriness, to handle numerous degradations intermingled in a single old photograph. Two branches are fused in the latent space, improving the ability to repair many flaws in antique photographs. The suggested solution outperforms cutting-edge techniques for restoring ancient images in terms of visual quality. Our project is better than the previous teams work by 10%.