Scholarlink

<p>In automotive body-in-white (BIW) production, the large number of weld spots and diverse defect types make traditional inspection methods inefficient and inaccurate. This paper proposes an improved lightweight weld defect detection model, DHN-YOLO. Built on the YOLOv11n framework, the model incorporates three enhancements: a Diverse Branch Block (DBB) in the backbone to enrich multi-scale feature representation; integration of a Hierarchical Spatial Attention Network (HSPAN) and a dynamic sampling module (DySample) in the neck to enhance sensitivity to small targets and edge details; and a lightweight Efficient Head (EH) to reduce redundancy and improve inference efficiency. Experiments on an eight-class weld defect dataset demonstrate that DHN-YOLO achieves 95.3% precision, 95.3% recall, and 97.1% mAP@0.5, outperforming mainstream lightweight models. Meanwhile, the parameter count and computational complexity are reduced to 1.71M and 4.2 GFLOPs. Deployment on an embedded platform further shows a 35.1% speed improvement compared with YOLOv11n, achieving better real-time performance and deployability. These results indicate that DHN-YOLO provides high accuracy with lower computational cost, offering an efficient and practical solution for weld quality inspection in intelligent manufacturing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DHN-YOLO: an improved lightweight method for weld spot detection in automotive body-in-white

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Saidong Yang,
  • Debo Liu,
  • Xingchong Li,
  • Shixiong Zhang,
  • Lu Mi

摘要

In automotive body-in-white (BIW) production, the large number of weld spots and diverse defect types make traditional inspection methods inefficient and inaccurate. This paper proposes an improved lightweight weld defect detection model, DHN-YOLO. Built on the YOLOv11n framework, the model incorporates three enhancements: a Diverse Branch Block (DBB) in the backbone to enrich multi-scale feature representation; integration of a Hierarchical Spatial Attention Network (HSPAN) and a dynamic sampling module (DySample) in the neck to enhance sensitivity to small targets and edge details; and a lightweight Efficient Head (EH) to reduce redundancy and improve inference efficiency. Experiments on an eight-class weld defect dataset demonstrate that DHN-YOLO achieves 95.3% precision, 95.3% recall, and 97.1% mAP@0.5, outperforming mainstream lightweight models. Meanwhile, the parameter count and computational complexity are reduced to 1.71M and 4.2 GFLOPs. Deployment on an embedded platform further shows a 35.1% speed improvement compared with YOLOv11n, achieving better real-time performance and deployability. These results indicate that DHN-YOLO provides high accuracy with lower computational cost, offering an efficient and practical solution for weld quality inspection in intelligent manufacturing.