SM-YOLO-HybridBridge: a lightweight deep learning model for surface defect detection and structural health monitoring in hybrid wooden–steel bridges

Abstract

Bridges are indispensable components of modern transportation infrastructure, yet conventional manual inspections remain time-consuming, subjective, and inefficient for large-scale monitoring. To overcome these limitations, this study proposes SM-YOLO-HybridBridge, a lightweight deep learning model designed for surface defect detection on hybrid wooden–steel railway bridges. The model accurately detects five representative defect types—Wood Crack (96.3%), Wood Decay (94.4%), Rivet Loss (99.2%), Bolt Loss (94.1%), and Steel Corrosion (87.8%)—demonstrating high precision across diverse visual conditions. The main contributions of this research are as follows. (1) We construct a rare and diverse dataset collected from a heritage-listed hybrid bridge, covering five annotated defect categories to support robust training and benchmarking. (2) We introduce the SM-YOLO framework, which integrates the simple attention module mechanism and minimum point distance intersection-over-union loss to enhance feature discrimination and geometric localization while maintaining computational efficiency suitable for unmanned aerial vehicle-based deployment. (3) We focus on the underexplored hybrid wooden–steel bridge category, offering new insights and technical pathways for mixed-material defect detection within structural health monitoring. Experimental results demonstrate that SM-YOLO-HybridBridge maintains high accuracy and lightweight efficiency on real engineering datasets, even under small-sample and low-contrast conditions. In contrast, newer and more complex models, such as YOLOv11, exhibit noticeably degraded performance in these practical scenarios, indicating that architectural complexity does not necessarily translate into better detection capability for field-acquired civil infrastructure images.