A Benchmark for Cultural-Heritage Deterioration Segmentation: Semi-Supervised Baselines, Auto-Discovery, and Calibration-Aware Evaluation

Cultural-heritage preservation increasingly relies on computing technologies for documentation, restoration, analysis, and risk-informed conservation. Pixel-level deterioration segmentation remains challenging because heritage archives are heterogeneous and weakly curated, expert annotations are scarce, foreground regions are highly imbalanced, and authentic damage differs substantially from standard computer-vision benchmarks. We address a practical archival setting in which many image fragments are available but only a small subset has reliable pixel-level masks. Our benchmark integrates archive auto-discovery, Mean Teacher semi-supervised learning, and calibration-aware evaluation within a reproducible protocol. On the curated MuralDH subset ( \(N=961\) ), we compare LR-ASPP, DeepLabV3+, U-Net++, and SegFormer under identical conditions. SegFormer (MiT-B2) provides the strongest discrimination–calibration trade-off on the held-out test set ( \(N=201\) ), achieving 94.9% pixel accuracy, 0.783 F1, 0.643 IoU, and 0.019 ECE. We also apply the protocol to DeepCrack ( \(N=537\) ) and CrackSeg9k ( \(N=9{,}159\) ) as auxiliary cross-domain stress tests. SegFormer remains the strongest model, while the lower CrackSeg9k performance exposes sensitivity to thin, elongated, low-contrast structures. Qualitative comparisons, error maps, MC-Dropout uncertainty maps, and failure analysis show how decorative contours, pigment transitions, and textured regions can be confused with deterioration. The benchmark therefore provides both segmentation performance and diagnostic evidence for interpreting model behavior in conservation-oriented workflows.