BCDC-YOLO: A Lightweight Detection Network for Anti-Corrosion Coating Surface Defects on Bridge Prestressed Corrugated Ducts

During the long-term service and construction of bridges, various types of surface defects frequently occur in the anti-corrosion coating on the surface of prestressed metallic corrugated ducts, severely threatening the overall durability of the bridge structure. To address the challenges of uneven optical imaging, high difficulty in coating defect detection, and low manual interpretation efficiency caused by the periodic linear groove structure of the corrugated ducts, this paper proposes the BCDC-YOLO (Bridge Corrugated Duct Coating defect detection–YOLO) method, which is tailored for inspecting surface defects in corrugated duct coatings. First, the captured images of coating surface defects are augmented and processed to construct a standardized dataset suitable for training deep learning models. Second, aiming at the limitations of the original YOLO network in detecting tiny and low-contrast defects within complex corrugated groove backgrounds, four improvement strategies are introduced to optimize the network model from both microscopic and macroscopic dimensions: deformable convolution, the contrast-enhanced adaptive SiLU (CE-ASiLU) activation function, the efficient channel attention (ECA) module, and an improved dual-index Gaussian Wasserstein distance loss function. Finally, the optimized network model is utilized to train and predict the surface coating defect dataset, and the identification outcomes are thoroughly compared and discussed with existing mainstream detection methods. The results demonstrate that the proposed method achieves a mean average precision (mAP) of 95.42 across all coating damages, with an inference speed of 66.8 FPS, which far surpasses the industrial real-time detection threshold. Notably, for hidden and extremely elusive tiny pitting defects, the single-class average precision (AP) yields an improvement of 12.46 compared with the baseline network. Ablation experiments confirm the effectiveness of the four introduced optimization strategies. Compared with other mainstream detection frameworks, the proposed BCDC-YOLO achieves the optimal balance between detection precision and real-time computational efficiency, thereby providing solid methodological and technical support for the intelligent inspection of surface coating defects.