Scholarlink

<p>Defect detection in photovoltaic systems is a key technology for intelligent operation and maintenance. However, traditional methods based on convolutional neural networks (CNNs) struggle to capture global context, resulting in insufficient feature representation. Furthermore, utilizing drone-collected data for precise real-time small object detection remains challenging. This paper proposes a novel model based on DETR (DEtection TRansformer)–the Small Object Enhancement Mechanism Detection Transformer (SOEM-DETR). This model introduces a multi-scale enhanced feature pyramid architecture: first, employing Spatial-to-Depth Convolution (SPD-Conv) to minimize information loss during shallow feature processing; second, efficiently fusing multi-scale features through the Cross-Stage Small Object Enhancement Module (CSP-SOEM); Finally, it employs wavelet resampling to more effectively preserve feature information. Experiments show that on the Photovoltaic Anomalies Detection dataset, compared with the baseline method, the proposed scheme improves the indicators mAP@50 and mAP@50:95 by 2.9% and 1.9%, respectively, and reduces the number of parameters by 2.5%. On the IRSTD-1K dataset, the Mean Average Precision (mAP@50) reaches 90.3%, which fully verifies the effectiveness of the proposed method in small target detection.</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

SOEM-DETR: an efficient real-time drone infrared image photovoltaic defect detection model

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Lei Peng,
  • Shao-Kai Zheng,
  • Dao-Lei Wang

摘要

Defect detection in photovoltaic systems is a key technology for intelligent operation and maintenance. However, traditional methods based on convolutional neural networks (CNNs) struggle to capture global context, resulting in insufficient feature representation. Furthermore, utilizing drone-collected data for precise real-time small object detection remains challenging. This paper proposes a novel model based on DETR (DEtection TRansformer)–the Small Object Enhancement Mechanism Detection Transformer (SOEM-DETR). This model introduces a multi-scale enhanced feature pyramid architecture: first, employing Spatial-to-Depth Convolution (SPD-Conv) to minimize information loss during shallow feature processing; second, efficiently fusing multi-scale features through the Cross-Stage Small Object Enhancement Module (CSP-SOEM); Finally, it employs wavelet resampling to more effectively preserve feature information. Experiments show that on the Photovoltaic Anomalies Detection dataset, compared with the baseline method, the proposed scheme improves the indicators mAP@50 and mAP@50:95 by 2.9% and 1.9%, respectively, and reduces the number of parameters by 2.5%. On the IRSTD-1K dataset, the Mean Average Precision (mAP@50) reaches 90.3%, which fully verifies the effectiveness of the proposed method in small target detection.