Restoration of damaged artworks based on a generative adversarial network

Ancient and contemporary artworks represent culture, heritage, and history. The artworks act as a bridge between the past and future of humankind. Preserving artwork is necessary for saving cultural heritage for future generations. However, artworks tend to deteriorate with time due to humidity, temperature, improper handling, and storage. Damages to artworks require unique restoration treatments, and, in many cases, they cannot be returned to their original state. Moreover, traditional artwork restoration methods are very costly and time-consuming. Further, the artwork restoration field is very complex as we not only require sharper and visually realistic restoration but also to preserve the artistic style and features of the artwork. The virtual restoration after digitizing the artworks can be very helpful in the actual restoration of the artworks. This paper works in this direction and proposes a novel generative adversarial network-based artwork restoration method that can digitally restore the damaged artwork, which can help in the physical restoration of the artworks. The proposed generative adversarial network uses a modified U-Net architecture for the generator part and pre-trained residual networks to build the generator's encoder part. Since the residual networks are pre-trained on millions of ImageNet images, they generate better feature embeddings out of the input artwork images, which in turn help in the generation of better-restored artwork images. The proposed artwork restoration method works in a single step and does not require the generation of masks. The proposed network has been evaluated on two different datasets using performance metrics such as peak signal-to-noise ratio, mean squared error, structural similarity index, and Frechet inception distance. Results indicate that the proposed network outperforms the existing artwork restoration methods.