Deep learning techniques for crack detection in St. Theodore Church, Cappadocia

Abstract

The preservation of cultural heritage structures requires continuous monitoring and proactive damage assessment as cultural heritage structures are highly vulnerable to environmental and anthropogenic factors. Although deep learning has shown promise for crack detection, current studies often rely on bounding boxes and lack the precision needed for heritage surfaces, which are characterized by complex geometries and textures. This study investigates the applicability of YOLO-based models for instance segmentation of cracks in cultural heritage structures, using Saint Theodore Church in Cappadocia, T & uuml;rkiye, as a case study. Saint Theodore Church, part of a UNESCO World Heritage site, exemplifies rock-hewn religious and historic architecture but has suffered significant deterioration over time. A novel domainspecific dataset, YTU-CrackIS, was developed, comprising high-resolution camera and UAV imagery to train and evaluate YOLOv8, YOLOv9 and YOLOv11 models. The study also investigates transfer learning using both COCO and general crack datasets to assess the influence of pre-training on performance. The results show that YOLOv8 outperforms other models, achieving a mean Average Precision (mAP) of 0.635 for detection and 0.553 for segmentation. Generalization tests on additional crack datasets from cultural heritage sites confirm that models trained on domain-specific data significantly outperform those trained on generic datasets. These findings emphasize the need for tailored deep learning approaches in cultural heritage preservation, where damage characteristics differ from conventional structural cracks. (c) 2026 Elsevier Masson SAS. All rights are reserved, including those for text and data mining, AI training, and similar technologies.