Multi-Granularity Recognition of Defects in Traditional Stone Carvings Under Small-Sample Conditions: A Data-Sufficiency Assessment for Condition Documentation

Condition documentation for traditional stone carvings requires comparable and repeatable records, yet field images are often limited by small samples, class imbalance, multi-defect co-occurrence, and diffuse defect boundaries. This study uses 643 field images from 26 courtyards in the Guanzhong region to examine classification granularity, model performance, and conservation applicability. Three tasks were defined: M1 for defect presence, M2 for six-class multi-label recognition, and M3 for eleven-class fine-grained recognition. Under courtyard-grouped nested cross-validation, in which the held-out fold is never used for model selection or threshold tuning, ResNet-50 achieved an F1 of 0.8908 ± 0.0055 on M1, with macro-F1 falling to 0.5378 ± 0.0148 on M2 and 0.3274 ± 0.0159 on M3. Model and training variants did not consistently outperform the baseline, and fine-tuned vision–language models reached comparable accuracy on M2 (0.5719) at far higher training cost. Class-wise analysis showed pronounced differences in reliability, with artificial damage at an F1 of 0.7361 and abrasion at 0.3409; false positives were more frequent in images containing other defect labels, while abrasion and soiling showed high missed-detection rates. M1 is suitable for defect screening and M2 can support defect typing and expert review, whereas M3 is presented as a data-sufficiency assessment.