Unsupervised domain-adaptive damage segmentation for Tibetan mural images

Abstract Tibetan murals are a cultural heritage form rich in historical and artistic information, but many surviving images are affected by cracks, flaking, pigment loss, abrasion, and other deterioration. Accurately segmenting damaged regions is challenging because of complex decorative textures, irregular degradation morphology, domain discrepancy, and limited pixel-level annotations. This study proposes an unsupervised domain-adaptive damage segmentation framework that transfers damage knowledge from labeled Dunhuang murals to unlabeled Tibetan mural images. The framework integrates DINOv3-based multi-layer visual representations, an FPN-inspired lightweight decoder, and teacher-student pseudo-label adaptation with dynamic filtering and target-domain reweighting. Experiments use 961 labeled Dunhuang mural image-mask pairs from MuralDH and 906 Tibetan mural images. On the held-out Tibetan test set, the method achieves 61.54% Dice and 44.45% IoU, supporting low-annotation damage localization for Tibetan mural preservation.