Scholarlink

<p>A novel surface detection model, designated as YOLO-DCFB-BiFPN-SENetV2 (YOLO-DBS), is introduced to address the limitations in detection accuracy of existing methodologies for surface defects in Small Out-line Package (SOP). Initially, the integration of the Coordinate Attention (CA) channel attention mechanism enhances the C2f module within the feature fusion network, thereby facilitating the successful identification of irregularly shaped defect features. Subsequently, to augment the model’s feature expression capability for minor defects while reducing its parameters, a lightweight and efficient multi-scale feature fusion network, termed BiFPN, is employed. Finally, the SENetV2 module is incorporated into the detection head to bolster the model’s ability to discern minor defects against similar backgrounds and mitigate external noise interference. Comparative experiments conducted on a self-constructed SOP dataset demonstrate that YOLO-DBS surpasses more advanced defect detection techniques, achieving a detection accuracy (mAP) of 99.4%. Moreover, the YOLO-DBS model’s parameter count is merely 2.5 million, which is 0.5 million fewer than that of the original model, illustrating how YOLO-DBS effectively balances model complexity and accuracy, thereby providing a reliable method for identifying surface defects in chip packaging within real-world industrial contexts.</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

YOLO-DBS: a multi-scale feature fusion-based surface defect detection method of small out-line package

  • 期刊论文

  • 发表日期：

    2025年9月7日

查看全文
  • Yikai Xing,
  • Xin Fang,
  • Yongbing Zhou,
  • Jian Zhang,
  • Haojie Chen

摘要

A novel surface detection model, designated as YOLO-DCFB-BiFPN-SENetV2 (YOLO-DBS), is introduced to address the limitations in detection accuracy of existing methodologies for surface defects in Small Out-line Package (SOP). Initially, the integration of the Coordinate Attention (CA) channel attention mechanism enhances the C2f module within the feature fusion network, thereby facilitating the successful identification of irregularly shaped defect features. Subsequently, to augment the model’s feature expression capability for minor defects while reducing its parameters, a lightweight and efficient multi-scale feature fusion network, termed BiFPN, is employed. Finally, the SENetV2 module is incorporated into the detection head to bolster the model’s ability to discern minor defects against similar backgrounds and mitigate external noise interference. Comparative experiments conducted on a self-constructed SOP dataset demonstrate that YOLO-DBS surpasses more advanced defect detection techniques, achieving a detection accuracy (mAP) of 99.4%. Moreover, the YOLO-DBS model’s parameter count is merely 2.5 million, which is 0.5 million fewer than that of the original model, illustrating how YOLO-DBS effectively balances model complexity and accuracy, thereby providing a reliable method for identifying surface defects in chip packaging within real-world industrial contexts.