SFMD-Net: mural damage prediction with semantic-frequency feature fusion and multi-scale structural decoding

Ancient mural damage is often small, fragmented, boundary-dominated, and visually similar to painted texture, complicating pixel-level condition assessment. We first characterise this challenge across mural, cultural-heritage, natural-image, and medical-image segmentation datasets. The analysis identifies YMDA as having a pronounced small-region burden, fragmentation, boundary dominance, and local texture overlap. We then propose SFMD-Net, a DINOv3-based framework combining a semantic–frequency interaction adaptor, high-fidelity projection bridge, and multi-scale damage-aware decoder to preserve weak damage cues while suppressing texture-induced responses. On YMDA, SFMD-Net achieves the highest Dice, intersection over union, precision, and recall among the compared models. On ARTeFACT, a heterogeneous analogue-media damage benchmark, it obtains the highest mean Dice and intersection over union under full target-domain training. YMDA-initialised fine-tuning further improves over target-only training. These results support SFMD-Net for in-domain mural damage prediction and supervised adaptation to heterogeneous heritage imagery, with potential value for condition assessment and assisted digital restoration.