A comprehensive review of deep learning methods in damage classification, detection, and segmentation of cultural heritage sites

Abstract

Deep learning techniques are increasingly used to monitor and assess damage in cultural heritage sites. This paper reviews recent advances in deep learning for classifying, detecting, and segmenting damage in the context of heritage preservation. Classification methods identify the type of damage (e.g., cracks, mould) but lack detailed spatial information. Detection methods use bounding boxes to localize damaged regions, thereby simplifying damage monitoring. Segmentation methods provide pixel-level mapping of damage

hence, they are useful for documenting complex structures and surfaces. However, all segmentation-based approaches require large datasets and computational resources. This review systematically compares these three methodologies, discussing the strengths and limitations of each with respect to dataset requirements, spatial precision, and computational demands. In addition, the application of hybrid models, transfer learning, and the combination of deep learning with traditional image processing methods are discussed in the context of cultural preservation. Based on this discussion, suitable approaches are suggested for different heritage monitoring tasks and scenarios. Furthermore, the paper outlines potential directions for further research. (c) 2026 Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.