AI-BASED CRACK DETECTION AND EVALUATION IN HISTORICAL BUILDINGS

Abstract

Artificial intelligence and computer vision techniques have emerged as powerful tools for automating crack detection in historical buildings. In the context of ensuring the structural integrity of historical masonry buildings, these techniques offer a robust solution. This task is important for safety and maintenance, where proactive detection of cracks can prevent damage and risks. However, detecting cracks within masonry structures remains a challenge. This thesis explores effective automated crack detection methods where three different computer vision approaches are studied. The first approach is focused on deploying a statistical-based model that depends on mathematical fixed formulations and local textural features to separate cracks from background. This approach is evolved into an enhanced ML-based model, where statistical and textural features are extracted to predict optimal detection thresholds. Finally, the third approach deployed deep learning models using transfer learning, leveraging pre-trained architectures to perform feature extraction. To support the data-intensive requirements of the deep learning model, the research involved extensive data collection conducted in the Old City of Nablus. Efforts were devoted to gathering data from multiple locations across the city to ensure diversity in structural conditions and crack patterns. In addition, pre-processing steps were implemented to standardize the dataset prior analysis. As a result, a dataset of 2794 images of masonry structures was constructed. All methods were evaluated on a consistent test set to ensure unbiased comparisons. The evaluation methodology was designed in collaboration with civil engineering experts to assess performance across two main dimensions: F1-Score, and detection rate. The F1-Score reflects the reliability and detection rate reflects safety assurance. The results demonstrated a clear hierarchy in performance. The statistical-based crack detection model achieved a baseline F1-Score of 0.4775. The ML-based model improved upon this with an F1-Score of 0.5151, confirming the advantage of data-driven parameter optimization. In the deep learning domain, the single class YOLOv8 model showed good reliability, achieving the highest F1-Score of 0.6116, effectively balancing precision with sensitivity. However, when evaluated on the detection rate, the single class Mask R-CNN proved superior performance, identifying 89.23% of all potential cracks.