CSEANet: Cross-Stage Enhanced Aggregation Network for Detecting Surface Bolt Defects in Railway Steel Truss Bridges

The accurate detection of surface bolt defects in railway steel truss bridges plays a vital role in maintaining structural integrity. Conventional manual inspection techniques require extensive labor and introduce subjective assessments, frequently yielding variable results across inspections. While UAV-based approaches have recently been developed, they still encounter significant technical obstacles, including small target recognition, background complexity, and computational limitations. To overcome these challenges, CSEANet is introduced-an improved YOLOv8-based framework tailored for bolt defect detection. Our approach introduces three innovations: (1) a sliding-window SAF preprocessing method that improves small target representation and reduces background noise, achieving a 0.404 mAP improvement compared with not using it; (2) a refined network architecture with BSBlock and MBConvBlock for efficient feature extraction with reduced redundancy; and (3) a novel BoltFusionFPN module to enhance multi-scale feature fusion. Experiments show that CSEANet achieves an mAP@50:95 of 0.952, confirming its suitability for UAV-based inspections in resource-constrained environments. This framework enables reliable, real-time bolt defect detection, supporting safer railway operations and infrastructure maintenance.