YOLODF: a concrete bridge surface damage detection model based on multiscale feature fusion in complex environments

Timely and efficient real-time surface damage detection is essential for maintaining the healthy operation of concrete bridges and has become a critical research focus. However, existing deep learning–based damage detection methods still face challenges such as low detection accuracy, poor adaptability, and limited applicability to diverse scenarios. To address these issues and enhance surface damage detection performance in complex environments, this study proposes an improved YOLODF model based on You Only Look Once, Version 5 (YOLOv5). The improvements include replacing the C3 module with the C2f structure with depthwise separable convolutions and inverted bottlenecks (DSIBC2f) module to build a new backbone network, DSIBCSPDarknet, which strengthens feature extraction capabilities. The SPPFCSPC structure is introduced to replace the spatial pyramid pooling fast (SPPF) module, enabling more effective multiscale feature fusion. Furthermore, the Enhanced Multidimensional Collaborative Attention (EMCA) is combined with the DSIBC2f module to construct a fused neck, FNeck, further optimizing feature fusion. Experimental results show that YOLODF significantly outperforms YOLOv5 in terms of precision, recall, F1 score, and mAP_0.5 and also surpasses the latest YOLOv12. Additionally, it demonstrates excellent damage detection capabilities in challenging scenarios, such as adverse weather, noise interference, and color variations. Despite a slight increase in computational load, YOLODF achieves a detection speed of 118 frames per second, demonstrating its high practicality for surface damage detection on bridges in complex environments.