Data-driven Detection and Evaluation of Damages in Concrete Structures: Using Deep Learning and Computer Vision

Abstract

Maintaining the structural integrity of concrete infrastructure—such as buildings, bridges, and tunnels—is critical to ensuring public safety and uninterrupted operations. However, traditional inspection methods remain labor-intensive, inconsistent, and impractical for large-scale deployment. This research investigates the application of cutting-edge, vision-based deep learning models to automate damage detection in concrete structures, with a focus on cracks and spalling. Two state-of-the-art architectures—YOLO-v7 instance segmentation and Mask R-CNN—are evaluated using an augmented dataset of 10,995 annotated images derived from an initial set of 400 real-world samples. Both models were trained via transfer learning on the COCO dataset and assessed using precision, recall, mean average Precision at IoU 0.5 (mAP50), and inference speed (FPS). YOLO-v7 achieved a superior mAP50 of 96.1% and a real-time processing rate of 40 FPS, making it ideal for rapid field deployment. In contrast, Mask R-CNN delivered a strong mAP50 of 92.1% at 18 FPS, favoring high segmentation fidelity for offline analysis. YOLO-v7’s Efficient Layer Aggregation Networks (E-ELAN) enable efficient real-time inference, while Mask R-CNN’s Region Proposal Network (RPN) enhances detailed damage localization. These findings suggest a dual-use framework: YOLO-v7 for proactive, on-site monitoring, and Mask R-CNN for post-event forensic evaluation. This study advances the integration of AI in structural health monitoring and paves the way for future research on hybrid architectures, broader damage typologies, and extension to other infrastructure domains, such as steel bridges and composite structures.