| description abstract | Fiber-reinforced polymers (FRPs) are indispensable in civil engineering owing to their high tensile strength, lightweight characteristics, and exceptional durability. Notably, carbon fiber-reinforced polymer (CFRP) concrete is distinguished by its superior mechanical properties and corrosion resistance. Despite these advantages, structural defects can arise at the CFRP-concrete interface, resulting in cracking and delamination that compromise structural integrity. Traditional defect detection methods encompass manual visual inspection and instrument-based detection utilizing physical signals. However, these approaches exhibit significant limitations in detection efficiency, identification accuracy, and cost-effectiveness. In light of this exigency, this study proposes a deep learning methodology for surface defect detection in CFRP concrete. This approach enhances the detection accuracy of the YOLOv8 model through the incorporation of a lightweight module (C2f-RVB-EMA) and the utilization of the Powerful-IoU loss function to compute overlap ratios, thereby augmenting the model’s generalization capabilities. The resultant model demonstrates commendable performance metrics including accuracy, recall, F1 score, mAP50, and mAP50-95, achieving values of 86.8%, 88.5%, 0.88, 87.9%, and 69.6%, respectively. Moreover, the compact size of the developed model, 6.2M, significantly mitigates computational overheads during both training and inference phases, rendering it amenable for deployment across various resource-constrained edge devices. | |
