Deep learning-based YOLO network model for detecting surface cracks during structural health monitoring

Abstract

Detecting surface cracks and identifying their severity are crucial steps in the structural health monitoring of historic buildings. In the field of heritage preservation, crack formation and other related defects are substantial flaws because they significantly affect the long-term durability of historic structures. In this study, an automated method based on the deep learning (DL) object detection model, You Only Look Once (YOLOv5), was deployed. It captures and pinpoints cracks in masonry structures via bounding boxes. The developed DL model was trained using 4000 annotated images collected using a mobile camera from different historic sites in Bhubaneswar, Odisha, India. The training time of the model was relatively short, which led to low computational costs of numerical simulations. The developed YOLOv5 model achieved a mean average precision (mAP_0.5) of approximately 92% on the collected masonry crack database. Unlike other contemporary DL models, this model can be used for real-time health monitoring. The findings of this study provide insights for identifying structural problems that require immediate repair, enabling improved monitoring and inspection of historic buildings.