Abstract:
This thesis addresses digital art restoration through the application of image inpainting, a technique for filling in missing sections of images while maintaining visual coherence. We analyze deep learning diffusion-based methodologies such as Latent Diffusion Model (LDM), EdgeConnect and ControlNet. These models are explored in this study due to their ability to generate high-quality images based on textual prompts and customized pixel-wise/based conditions. The primary aim is to utilize these models to devise a novel approach for enhancing the restoration of frescoes, and to evaluate it on the case study of the RePAIR Project, which aims to revolutionize archaeology through the integration of computer vision, and artificial intelligence for the restoration of fragmented artifacts.
The thesis delves into an exploration of diffusion models, analyzing their mathematical foundations and basic architectures, with particular attention to recent advancements like the LDM. It also discusses the Image Inpainting task, presenting key architectures such as LDM, EdgeConnect, and ControlNet, while examining various methodologies and experimental setups from related literature to offer a comparative analysis.
