YOLO-WP: A Lightweight and Efficient Algorithm for Small-Target Detection in Weld Seams of Small-Diameter Stainless Steel Pipes

To address the low detection efficiency and high computational resource demands of current welded pipe defect detection algorithms for small target defects, this paper proposes the YOLO-WP algorithm based on YOLOv5s. The improvements of YOLO-WP are mainly reflected in the following aspects: First, an innovative GhostFusion architecture is introduced in the backbone network. By replacing the C3 modules with C2f modules and integrating the Ghost CBS module inspired by Ghost convolution, cross-stage feature fusion is achieved, significantly enhancing computational efficiency and feature representation for small target defects. Second, the Slim-Neck lightweight design based on GSConv is employed in the neck to further optimize the network structure and reduce the number of parameters. Additionally, the SimAM lightweight attention mechanism is incorporated to improve the network's ability to extract defect features, and the Focal-EIou loss is utilized to optimize CIou loss, thereby enhancing small object detection and accelerating loss convergence. The experimental results show that the AP(D1) and mAP@0.5 of the YOLO-WP model are improved by 5.3% and 3%, respectively, over the original model. In addition, the number of model parameters and FLOPs are reduced by 40% and 45%, respectively, achieving a good balance between performance and efficiency. We evaluated the performance of YOLO-WP using other datasets and showed that YOLO-WP exhibits excellent applicability. Compared to existing mainstream detection algorithms, YOLO-WP is more advanced. The YOLO-WP model significantly enhances production quality in industrial defect detection, laying the foundation for building compact, high-performance embedded weld pipe surface defect detection systems.