MuHP-DLNet: a multi-stage high-precision mural damage detection network with damage-aware dynamic loss

Ancient murals exhibit sparse and irregular damage, creating severe foreground imbalance and challenging boundary segmentation. We propose MuHP-DLNet, a multistage network combining a coarse branch for global localization with a fine branch for detailed segmentation. The fine branch integrates edge enhancement, Transformer encoding, and query guided dynamic feature modulation. Cross scale aggregation further combines global context with local details to improve mask coherence and boundary delineation. A damage aware loss combines weighted binary cross entropy and Dice using image damage ratios and local annotation structure, while learnable normalized weights balance main and auxiliary supervision. Experiments on two mural datasets and three public benchmarks demonstrate favorable overall performance against representative baselines. Controlled ablations identify complementary architectural contributions and show that the proposed loss has the largest effect among the evaluated components. These results support the method’s effectiveness for segmenting subtle damage within complex mural textures.