An Improved YOLOv9‐Based Lightweight Method for Surface Defect Detection in Weld Seams

ABSTRACT

To address issues such as insufficient feature representation capability for small‐sized defects and significant interference from complex backgrounds in weld surface defect detection, this paper proposes a lightweight weld defect detection network, CS‐YOLO (You Only Look Once). First, we construct a cross‐scale feature fusion pyramid network to enhance defect feature representation through multi‐scale feature interaction. Second, we design a spatial‐to‐depth transformation convolution module to improve the representation capability of small‐object features by reducing the loss of fine‐grained information during downsampling. Finally, we introduce the normalised Wasserstein distance to replace IoU as the bounding box similarity metric, thereby improving regression stability in small‐object detection scenarios. Experimental results on the self‐built WELD‐DETECT dataset show that, compared with the baseline model YOLOv9, CS‐YOLO achieves a 2.4% improvement in mAP@50 and a 4.2% increase in Precision, without introducing additional parameter overhead. External validation on the public LoHi‐WELD dataset demonstrates that CS‐YOLO maintains stable detection performance across different weld defect detection scenarios, verifying its generalisation capability.