Scholarlink

<p>Existing object detection algorithms for insulator defect detection in complex backgrounds face critical challenges. These include missed detections, false positives, low detection efficiency, and excessive model sizes that impede edge device deployment. To address these issues, this paper presents the YOLOv5-lite, a lightweight insulator defect detection network improved from YOLOv5. In the feature extraction backbone, depthwise separable convolutions are employed to improve CSPDarknet, which reduces model parameters and enhances detection speed. For the feature fusion stage, a hybrid attention mechanism is proposed. This mechanism integrates coordinate attention and multi-head self-attention to resolve shallow feature loss in the feature pyramid network, thereby improving the model’s detection accuracy. The loss function is optimized through integration of Focal loss and EIOU. This improvement mitigates the imbalance between positive and negative samples and accelerates convergence of the training model. Significant performance gains are validated by the experimental results. Compared with the baseline YOLOv5 model, the proposed YOLOv5-lite achieves a 9.23% increase in detection accuracy, with a mean average precision (mAP) of 97.05%. Meanwhile, it reduces model parameters by 52.5% and boosts detection speed by 77.41% to 91.9 frames per second. These results satisfy the requirements for high-precision real-time power inspection applications.</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

YOLOv5-lite: A Highly Accurate Insulator Defect Detection Algorithm Based on Hybrid Attention

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Honglin Nie,
  • Yong Li,
  • Sheng Han

摘要

Existing object detection algorithms for insulator defect detection in complex backgrounds face critical challenges. These include missed detections, false positives, low detection efficiency, and excessive model sizes that impede edge device deployment. To address these issues, this paper presents the YOLOv5-lite, a lightweight insulator defect detection network improved from YOLOv5. In the feature extraction backbone, depthwise separable convolutions are employed to improve CSPDarknet, which reduces model parameters and enhances detection speed. For the feature fusion stage, a hybrid attention mechanism is proposed. This mechanism integrates coordinate attention and multi-head self-attention to resolve shallow feature loss in the feature pyramid network, thereby improving the model’s detection accuracy. The loss function is optimized through integration of Focal loss and EIOU. This improvement mitigates the imbalance between positive and negative samples and accelerates convergence of the training model. Significant performance gains are validated by the experimental results. Compared with the baseline YOLOv5 model, the proposed YOLOv5-lite achieves a 9.23% increase in detection accuracy, with a mean average precision (mAP) of 97.05%. Meanwhile, it reduces model parameters by 52.5% and boosts detection speed by 77.41% to 91.9 frames per second. These results satisfy the requirements for high-precision real-time power inspection applications.