This master's thesis explores the use of deep learning for the reconstruction of missing parts of archaeological artifacts, focusing on 2D images of frescoes and mosaics. Traditional restoration methods are time-consuming, costly, and often irreversible, prompting an investigation into automated approaches using artificial intelligence, specifically the Stable Diffusion XL (SDXL) model. The base model was fine-tuned using the LoRA technique to capture the style of the same artist or similar works and applied to case studies involving frescoes from the Church of the Virgin Mary in Polje, the Villa of the Mysteries in Pompeii, the Franciscan Church in Ljubljana, and a mosaic from Mošnje. The reconstructions were evaluated using quantitative metrics SSIM and LPIPS, as well as qualitative assessments by an expert. The results demonstrate that the reconstructions are visually compelling but require additional understanding of iconography and context to ensure historical accuracy. The proposed approach offers a faster and more cost-effective alternative to traditional methods, though it necessitates further improvements in incorporating iconographic references for ethical and precise restoration.
|
