WDI-YOLO: A lightweight steel bridge weld defect detection algorithm using UAV images
Wei Ji, Shibo Liu, Lu Deng, J. Li, Yang Liu, Zhi Xiong
Abstract
Ensuring weld integrity is essential for the durability and safety of steel bridges, yet inspection in elevated bridge structures or complex environments remains difficult. This study proposes WDI-You Only Look Once (YOLO), a lightweight weld defect detection algorithm designed for unmanned aerial vehicle (UAV)-based inspection, built upon the YOLOv8-n framework. The model is trained on a dataset comprising 4024 augmented images derived from original weld images captured by UAVs and in factory environments. To enhance feature representation and reduce computational overhead, standard convolution modules in the Backbone are replaced with Adaptive Downsampling (ADown) and Grouped Shuffle Convolution (GSConv). The Spatial Pyramid Pooling-Fast (SPPF) module is substituted with the Spatial Pyramid Pooling with Enhanced Local Attention Network (SPPELAN) to improve the detection of small-scale defects. A Spatial and Channel Synergistic Attention (SCSA) mechanism is embedded into the C2f module, with the output channels of the 23rd layer halved for further model compression. Triplet Attention is introduced before the detection head to mitigate background interference. Additionally, Quality Focal Loss is employed to improve localization accuracy and generalization, particularly for small targets. Experimental results show a 5.5 % increase in [email protected] and a 21.6 % reduction in parameters compared to the baseline YOLOv8-n. WDI-YOLO also surpasses other mainstream detection algorithms, achieving 93.7 % precision with only 7.1 GB of FLOPs. The lightweight design of WDI-YOLO enables efficient and accurate automated weld defect detection on UAVs, significantly enhancing safety and reducing maintenance costs for steel bridges.