Style-Infused Image Restoration with GAN
Jingzhi Gao, Xuesong Su · 2023
The restoration of old and damaged photographs has consistently been a popular issue. Old photos serve as records of bygone, cherished moments in life, carrying a myriad of emotions of the individuals who took them. However, due to limitations in photographic conditions at the time or improper preservation, these photos often exhibit mold spots, dirt, and scratches. Therefore, the restoration of old photographs holds significant importance. Conventional restoration methods primarily rely on manual intervention and typically yield mediocre results. We propose a novel two-stage network structure. In the first stage, it detects and labels scratches and dirt in old or damaged photos, forming a mask. In the second stage, a GAN network is used to generate the missing portions based on the mask. Importantly, we embed latent encoding into the network to fine-tune the image style, thereby enhancing the restoration outcome. Extensive experiments have demonstrated that our network achieves promising results in the task of restoring old and damaged photos.