Scholarlink

<p>Detecting whether underground workers are wearing safety helmets has long been an essential yet challenging task in industrial safety monitoring. Due to dim lighting, occlusion, and complex backgrounds in mining environments, existing object detection models often fail to achieve satisfactory performance. To address these issues, this study proposes an improved helmet detection approach based on YOLOv5. The Squeeze-and-Excitation Network (SENet) is enhanced with a multi-scale average pooling mechanism and the smoother Swish activation function, forming a lightweight attention module named SENet-tn. In addition, a novel loss function, INWDLoss, is introduced by integrating the advantages of the Normalized Wasserstein Distance (NWDLoss) and IoULoss to improve robustness in small-target detection and high-overlap scenarios. Experiments conducted on the self-built Mine Wearing Helmet Dataset (MWHD) demonstrate that the proposed YOLOv5s-tn model increases the mean Average Precision (mAP) by 2.91% compared to the baseline YOLOv5s. These results confirm that the enhanced attention mechanism and loss function effectively improve detection accuracy and reliability for underground helmet-wearing scenarios, providing technical support for intelligent safety supervision in mines.</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

Research on attention mechanism and loss function based on YOLOv5 to improve the detection method of underground personnel wearing safety helmets

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月31日

查看全文
  • Gengmu Qin,
  • Jingtao Pan,
  • Dan Zhao,
  • Chuanxu Zhang

摘要

Detecting whether underground workers are wearing safety helmets has long been an essential yet challenging task in industrial safety monitoring. Due to dim lighting, occlusion, and complex backgrounds in mining environments, existing object detection models often fail to achieve satisfactory performance. To address these issues, this study proposes an improved helmet detection approach based on YOLOv5. The Squeeze-and-Excitation Network (SENet) is enhanced with a multi-scale average pooling mechanism and the smoother Swish activation function, forming a lightweight attention module named SENet-tn. In addition, a novel loss function, INWDLoss, is introduced by integrating the advantages of the Normalized Wasserstein Distance (NWDLoss) and IoULoss to improve robustness in small-target detection and high-overlap scenarios. Experiments conducted on the self-built Mine Wearing Helmet Dataset (MWHD) demonstrate that the proposed YOLOv5s-tn model increases the mean Average Precision (mAP) by 2.91% compared to the baseline YOLOv5s. These results confirm that the enhanced attention mechanism and loss function effectively improve detection accuracy and reliability for underground helmet-wearing scenarios, providing technical support for intelligent safety supervision in mines.