YOLO-MAFD: A Collaborative Detection Framework for Automated Recognition of Bridge Steel Structural Components

Automatic recognition of bridge steel structural components is crucial for intelligent bridge inspection and structural assessment. This study addresses a four-class detection task involving bearings, out-of-plane stiffeners (OPS), gusset plate connections (GPC), and cover plate terminations (CPT) in complex bridge scenes. Although YOLOv8 provides an efficient detection framework, its performance is limited by insufficient discriminative feature extraction, information loss during multi-scale fusion, and instability in classification and localization under background interference and class similarity. To address these limitations, a collaborative detection architecture, termed YOLO-MAFD, is proposed by introducing an attention-enhanced convolution module into the backbone, an optimized multi-scale feature fusion strategy into the neck, and an improved decoupled prediction head. Experimental results show that the proposed method outperforms the baseline YOLOv8 model, achieving an mAP@0.5 of 0.766, corresponding to a 9.4% relative improvement, while also improving mAP@0.5:0.95 and Recall. These results demonstrate that the proposed method improves the accuracy and robustness of bridge steel structural component detection, particularly under complex backgrounds and multi-class interference.