Scholarlink

<p>With the widespread adoption of satellite remote sensing, detecting overhead line towers from satellite imagery has become a research hotspot. However, existing detectors for optical satellite imagery lack mechanisms that simultaneously model multi-scale backgrounds and preserve small, edge-like tower structures. To address this limitation, this article introduces MD-YOLO (mind detail–you only look once), an enhanced target detection algorithm used to improve the detection accuracy of overhead transmission line towers in satellite remote sensing images. MD-YOLO integrates a large selective kernel network (LSKNet) and space-to-depth cross stage partial network (SPDCSPNet). LSKNet improves multi-scale feature extraction, mean average precision (mAP), and recall, especially in multi-scale scenarios. SPDCSPNet optimizes the neck structure, enhancing robustness in complex backgrounds and occlusions. Additionally, replacing standard convolutions in the detection head with SPD-Conv improves small-target representation. Experimental results show MD-YOLO achieves 2.2% higher mAP@0.5, 7.4% higher mAP@0.5:0.95, and a 3.7% increase in recall compared to the YOLOv8n baseline on the ZJKX dataset. Moreover, evaluations on the VisDrone2019 dataset further confirm MD-YOLO’s generalization ability, maintaining competitive accuracy with lower complexity. These results demonstrate MD-YOLO’s efficiency and applicability for intelligent overhead line tower inspection.</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 with multi-scale optimization for overhead line tower detection in satellite remote sensing images

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Shouming Hou,
  • Boshu Wang,
  • Ning Wang,
  • Jianchao Hou,
  • Yuteng Pang

摘要

With the widespread adoption of satellite remote sensing, detecting overhead line towers from satellite imagery has become a research hotspot. However, existing detectors for optical satellite imagery lack mechanisms that simultaneously model multi-scale backgrounds and preserve small, edge-like tower structures. To address this limitation, this article introduces MD-YOLO (mind detail–you only look once), an enhanced target detection algorithm used to improve the detection accuracy of overhead transmission line towers in satellite remote sensing images. MD-YOLO integrates a large selective kernel network (LSKNet) and space-to-depth cross stage partial network (SPDCSPNet). LSKNet improves multi-scale feature extraction, mean average precision (mAP), and recall, especially in multi-scale scenarios. SPDCSPNet optimizes the neck structure, enhancing robustness in complex backgrounds and occlusions. Additionally, replacing standard convolutions in the detection head with SPD-Conv improves small-target representation. Experimental results show MD-YOLO achieves 2.2% higher mAP@0.5, 7.4% higher mAP@0.5:0.95, and a 3.7% increase in recall compared to the YOLOv8n baseline on the ZJKX dataset. Moreover, evaluations on the VisDrone2019 dataset further confirm MD-YOLO’s generalization ability, maintaining competitive accuracy with lower complexity. These results demonstrate MD-YOLO’s efficiency and applicability for intelligent overhead line tower inspection.