Crack detection in churches using deep learning

Abstract Post-earthquake damage assessment of churches is a critical task, due to the architectural complexity and the safety risks associated with on-site inspections, which expose surveyors to potentially dangerous structurally compromised areas. Among visible indicators, cracks represent one of the most relevant features for evaluating structural damage. As part of the wider project RAISE (NRRP-ECS00000035), this study proposes an image-based deep learning approach for automatic crack detection in church buildings, with the aim of supporting traditional survey procedures and reducing both inspection time and operator risk. A convolutional neural network based on transfer learning is developed and evaluated using the Xception architecture. The model is first trained on a large benchmark dataset already available of concrete crack images. Image patches are extracted and classified in a binary framework (crack vs. non-crack), and model performance is assessed using standard metrics such as Precision, Recall, and F1-score. While initial training on concrete images yields satisfactory results (98% accuracy), performance drops significantly when applied to church images, highlighting a strong domain shift. Retraining with church-specific data, a custom dataset collected from churches in the Liguria region, including surfaces characterized by frescoes, plaster, and decorative elements, substantially improves performance. It achieves a Precision of 93.87%, a Recall of 88.45%, and an F1-score of 91.08% on frescoed surfaces of the vaults, a fundamental element in the damage survey phase. Additional tests investigate the influence of lighting conditions, image resolution, and capture distance, providing practical guidance for acquisition protocol design. The results demonstrate the potential of deep learning techniques for supporting crack detection in heritage contexts, while also highlighting current limitations related to data variability and patch-level classification. Future developments will focus on dataset expansion, pixel-level segmentation approaches, and integration with standardized damage assessment frameworks.