Unpaired Image-to-Image Translation Based Artwork Restoration Using Generative Adversarial Networks

Artworks assume huge significance in humankind’s culture and history. Artworks assimilate history and have the precise ability to tell history to the future generation. However, as time passes, environmental factors severely affect artworks, and these damages are often complicated to repair manually and through traditional methods. Conventional artwork restoration methods can not conserve the artist’s style and the artwork's semantic details. Restoration of the artworks through virtual restoration methods can be done using generative adversarial networks. After that, the virtually restored artwork can guide the physical restoration task of artworks. However, most of the methods based on generative adversarial network performs paired image-to-image translation for restoration, which necessitates the presence of the original version of the damaged artwork. This work proposes an unpaired image-to-image translation-based artwork restoration method using cycle-consistent generative adversarial networks. The proposed method does not require the availability of the original version of each damaged artwork for restoration. The network has been trained on a publicly available dataset of artwork images. The experimental results have been compared with other methods used for artwork restoration, and quantitative results demonstrate that the proposed approach performs better than the current and previous methods for the restoration of damaged artwork.