Scholarlink

<p>Photovoltaic (PV) panels play a crucial role in solar power systems, and their proper functioning is essential for overall system efficiency. Conventional inspection approaches struggle with large-scale PV arrays due to their inability to cope with complex backgrounds, significant scale variations, and dense panel distributions. Furthermore, intricate environmental interferences, heterogeneous panel appearances, and heavy occlusions exacerbate these challenges. To this end, we propose YOLO-PPM, a lightweight detection network designed specifically for PV panels. To manage scale variations and accentuate subtle defects in homogeneous backgrounds, and enhance global and local contextual understanding, our method introduces a Pixel-Guided Spatial Aggregation (PGSA) module that integrates Inception Depthwise Channel Group (IDCG) and Hybrid-Spatial Aggregation Strategies (HSAG). To mitigate the loss of multi-scale object information in deeper layers, we incorporate a Haar-based Transform Multi-Resolution Feature Pyramid Network (HTMFPN). Finally, to further refine defect detection in PV-specific scenarios, a Dynamic Hard-Sample Re-weighting (DHSR) strategy is employed. Extensive experiments demonstrate that the proposed lightweight YOLO-PPM achieves mAP.5 of 82.1% and mAP.5:.95 of 57.7%, while specifically improving detection for heavily occluded targets with mAP.5 of 47.5%. All these results are achieved with an efficient design of only 2.68M parameters, 8.7G FLOPs, and 5.8MB storage. Source code is available at <a href="https://github.com/SCNU-RISLAB/YOLO-PPM.">https://github.com/SCNU-RISLAB/YOLO-PPM.</a></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-PPM: a lightweight object detector with multi-frequency aware for photovoltaic panel matrix defect detection

  • 期刊论文

  • 发表日期：

    2026年2月22日

查看全文
  • Yunxing Zhu,
  • Jianping Yue,
  • Yating Li,
  • Bohuan Xue,
  • Xiaoyu Tang

摘要

Photovoltaic (PV) panels play a crucial role in solar power systems, and their proper functioning is essential for overall system efficiency. Conventional inspection approaches struggle with large-scale PV arrays due to their inability to cope with complex backgrounds, significant scale variations, and dense panel distributions. Furthermore, intricate environmental interferences, heterogeneous panel appearances, and heavy occlusions exacerbate these challenges. To this end, we propose YOLO-PPM, a lightweight detection network designed specifically for PV panels. To manage scale variations and accentuate subtle defects in homogeneous backgrounds, and enhance global and local contextual understanding, our method introduces a Pixel-Guided Spatial Aggregation (PGSA) module that integrates Inception Depthwise Channel Group (IDCG) and Hybrid-Spatial Aggregation Strategies (HSAG). To mitigate the loss of multi-scale object information in deeper layers, we incorporate a Haar-based Transform Multi-Resolution Feature Pyramid Network (HTMFPN). Finally, to further refine defect detection in PV-specific scenarios, a Dynamic Hard-Sample Re-weighting (DHSR) strategy is employed. Extensive experiments demonstrate that the proposed lightweight YOLO-PPM achieves mAP.5 of 82.1% and mAP.5:.95 of 57.7%, while specifically improving detection for heavily occluded targets with mAP.5 of 47.5%. All these results are achieved with an efficient design of only 2.68M parameters, 8.7G FLOPs, and 5.8MB storage. Source code is available at https://github.com/SCNU-RISLAB/YOLO-PPM.