Scholarlink

<p>Accurate powerline detection is crucial for ensuring stable power supply, but the detection accuracy of common models is often compromised in complex backgrounds, particularly when the targets are concealed. In this study, an adaptive multi-scale information fusion detection module based on YOLO is proposed, which identifies several objects that may cause interference to powerlines through a visual detection model. The Residual 2 Network (Res2Net) is proposed as the bottleneck structure of the Cross-Stage Partial Network (C3) module to enhance the network’s capability for multi-scale feature representation. Moreover, the ECA is used to facilitate the multi-scale information fusion at the feature output stage. Additionally, fixed convolution block is replaced with deformable convolution block to improve small target detection. The model is compressed with the adaptive Layer-adaptive Sparsity for the Magnitude-based Pruning (LAMP) pruning method so as to significantly reduce the number of parameters. Experimental results demonstrate that different models of YOLOv5 and YOLOv8, with an adaptive multi-scale information fusion detection module (C3_RED) improvement, exhibit superior performance in powerline detection. Post-compression, the model still achieves a high detection accuracy of 90.5% and operates at 262.8 FPS, thereby satisfying the requirements for real-time powerline detection. Compared with the benchmark model, mAP has increased by 6.69%. Furthermore, through LAMP pruning, we reduced the number of model parameters by 83.5% while maintaining high precision, demonstrating the great potential of C3_RED in real-time monitoring of edge devices.</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

A Multi-Scale Improved Fusion Network for Powerline Target Detection

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Binglin Li,
  • Jun Li,
  • Jiacheng Tang,
  • Lixin Wang

摘要

Accurate powerline detection is crucial for ensuring stable power supply, but the detection accuracy of common models is often compromised in complex backgrounds, particularly when the targets are concealed. In this study, an adaptive multi-scale information fusion detection module based on YOLO is proposed, which identifies several objects that may cause interference to powerlines through a visual detection model. The Residual 2 Network (Res2Net) is proposed as the bottleneck structure of the Cross-Stage Partial Network (C3) module to enhance the network’s capability for multi-scale feature representation. Moreover, the ECA is used to facilitate the multi-scale information fusion at the feature output stage. Additionally, fixed convolution block is replaced with deformable convolution block to improve small target detection. The model is compressed with the adaptive Layer-adaptive Sparsity for the Magnitude-based Pruning (LAMP) pruning method so as to significantly reduce the number of parameters. Experimental results demonstrate that different models of YOLOv5 and YOLOv8, with an adaptive multi-scale information fusion detection module (C3_RED) improvement, exhibit superior performance in powerline detection. Post-compression, the model still achieves a high detection accuracy of 90.5% and operates at 262.8 FPS, thereby satisfying the requirements for real-time powerline detection. Compared with the benchmark model, mAP has increased by 6.69%. Furthermore, through LAMP pruning, we reduced the number of model parameters by 83.5% while maintaining high precision, demonstrating the great potential of C3_RED in real-time monitoring of edge devices.