SDC-YOLO for proactive defense against welding defects: a lightweight real-time detection model integrating geometric perception and dynamic sampling

The welding quality of heavy energy equipment is a core element in ensuring structural integrity. Traditional non-destructive testing and evaluation (NDT&E) largely relies on offline defect detection after welding, making it difficult to meet the online quality monitoring needs during the manufacturing process. The development of quantitative non-destructive evaluation (QNDE), on the other hand, places higher requirements on the accurate perception of the spatial and geometric characteristics of surface defects. At the same time, existing detection models struggle to balance performance and real-time efficiency when screening for low-contrast and elongated anisotropic defects online. This paper proposes a lightweight in-situ weld defect detection model, SDC-YOLO. A Strip Attention Downsampling (SAD) module is designed for slender targets, a Dynamic Sampling (DySample) module is introduced for edge alignment, and a C3k2_PCA backbone is constructed to suppress redundancy. It has 2.50M parameters and 6.0 GFLOPs, achieving 160 FPS and 89.3% mAP@0.5 on a self-built weld defect dataset (2.3% higher than YOLO11n). LoHi-WELD tests confirm its generalisation.The model can stably output compact defect spatial candidate boxes (ROI) with low redundancy, providing an efficient front-end visual perception solution for in-situ quality monitoring and subsequent geometric quantitative evaluation in the welding manufacturing process.