Scholarlink

<p>Defect detection in industrial printed circuit boards is critical for ensuring the performance and reliability of electronic products. However, real-world PCB defects are often small, irregular, low-contrast, and appear in cluttered backgrounds, making accurate and efficient detection highly challenging. Existing models either emphasize local feature learning but lack global semantic awareness, or introduce costly global modeling at the expense of real-time performance. To tackle these limitations, we propose TPVG-YOLO, a novel lightweight detection framework that achieves a high balance between accuracy and efficiency. It incorporates two key modules: Twined-Path Convolution, a dual-branch structure that enhances spatial detail and structural consistency with minimal overhead; and Vision-Gated Fusion, a Group-mLSTM-based module that efficiently encodes directional global context without attention overhead. Together, they enable the model to capture fine-grained local features and long-range dependencies while maintaining a compact design. Extensive experiments on PKU-Market-PCB and RealWorld (RW-PCB) datasets demonstrate that TPVG-YOLO achieves mAP50 scores of 89.3% and 87.2%, respectively, outperforming YOLOv5, YOLOv8, YOLOv9, and YOLOv10, and approaching YOLOv11 and YOLOv12 accuracy with only 2.0M parameters and 5.2 GFLOPs. Our code will be available at <a href="https://github.com/JEFfersusu/TPVG-YOLO">https://github.com/JEFfersusu/TPVG-YOLO</a>.</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

TPVG-YOLO: Twined-Path Convolution and Vision-Gated Fusion for Efficient PCB Defect Detection

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Weichao Pan,
  • Xu Wang,
  • Chengze Lü

摘要

Defect detection in industrial printed circuit boards is critical for ensuring the performance and reliability of electronic products. However, real-world PCB defects are often small, irregular, low-contrast, and appear in cluttered backgrounds, making accurate and efficient detection highly challenging. Existing models either emphasize local feature learning but lack global semantic awareness, or introduce costly global modeling at the expense of real-time performance. To tackle these limitations, we propose TPVG-YOLO, a novel lightweight detection framework that achieves a high balance between accuracy and efficiency. It incorporates two key modules: Twined-Path Convolution, a dual-branch structure that enhances spatial detail and structural consistency with minimal overhead; and Vision-Gated Fusion, a Group-mLSTM-based module that efficiently encodes directional global context without attention overhead. Together, they enable the model to capture fine-grained local features and long-range dependencies while maintaining a compact design. Extensive experiments on PKU-Market-PCB and RealWorld (RW-PCB) datasets demonstrate that TPVG-YOLO achieves mAP50 scores of 89.3% and 87.2%, respectively, outperforming YOLOv5, YOLOv8, YOLOv9, and YOLOv10, and approaching YOLOv11 and YOLOv12 accuracy with only 2.0M parameters and 5.2 GFLOPs. Our code will be available at https://github.com/JEFfersusu/TPVG-YOLO.