Industrial welding often exhibits some essential problems, such as unclear defect characteristics and complex background information. However, the existing defect detection models have relatively high costs and may be weak in weld surface defect detection. To address the problem, this paper proposes an improved lightweight YOLOv11 model for welding surface defect detection, called YOLO-Air. First, the model integrates the feature extraction module with the convolutional module to boost feature representation capability and optimize computational efficiency. Second, the GSConv and VOV-GSCSP modules are embedded in the neck network to reduce feature redundancy of spatial and channel dimensions, and then lower the computational load. Third, a lightweight detection head is designed as part of the detection network to further reduce model complexity. Lastly, we compare our proposed YOLO-Air model with the baseline on the Welding Defect Test-V2 and NEU-DET datasets. Experimental results demonstrate that the proposed model yields superior performance for weld surface defect detection. Specifically, it improves the mAP50 metric by 1.3%, while reducing the number of parameters by 17.3% and the computational complexity by 31.7%.All key experimental data have passed robustness tests.
Zhang et al. (2026) studied this question.