Deep Latent Space Translation for Restoration and Enhancement of Vintage Photographs

In this study, A scheme is proposed ahead of machine study to turn back to an old photo in a severe. This is a good idea since traditional recovery methods cannot recover the genuine defects due to the wide range of degradation and domain gaps between simulated and real pictures. Concurrently, a new triplet domain transformation network is made, using both real photos and synthetic image pairs. The network consists of two variational autoencoders (VAEs) one is learned by teaching it the old, clean photos separately from the other, then the other one and the translation between these spaces using synthetic paired data. This method of teaching closes the gap and becomes a good method for real-life photos. The approach also involves a global branch for structured defects, located in one location and a local branch for the unstructured defects fused on another spot which later is added to latent space is an approach that allows improving of the restoration quality. Furthermore, a facial refining network is used to recover fine facial details that will lead to the production of high-quality images. The translation between these latent spaces, facilitated by synthetic paired data, helps bridge the domain gap, improving the restoration of real-world images.