Scholarlink

<p>Electroluminescence (EL) imaging is currently the most efficient method for defect detection in photovoltaic modules. However, existing methods still have limitations in addressing complex industrial scenarios, failing to meet high-precision and real-time quality inspection requirements. This paper proposes a novel dual-branch multi-scale aggregation detection method based on improved YOLOv8, aiming to solve three key challenges in EL image defect detection: shape variations of homologous defects, polysilicon substrate background complexity (including overlapping defect identification), and high local detail similarity among different defect types. The framework has three core innovations: (1) its dual-branch backbone uses a parallel coarse-grained branch to capture global features of large-area defects and a fine-grained branch to focus on local details of tiny defects; their complementarity resolves feature adaptation issues caused by homologous defect shape differences; (2) the lightweight coordinate attention fusion module integrates spatial coordinate encoding into attention mechanisms to accurately locate defects in complex backgrounds and reduce polysilicon substrate background texture interference; (3) the designed cross-spatial feature pooling module enhances multi-scale feature distinguishability and captures subtle differences between similar defects via parallel multi-round max pooling and efficient multi-scale attention (EMA), improving the model’s classification accuracy for similar defects. Integrating these modules with an optimized YOLOv11 detection head, the system performs outstandingly on the PVEL-AD-2021 and PV-Multi-Defect datasets: mAP50 scores reach 90.1% and 83.8% (+&#xa0;8.7%, +&#xa0;5.1% vs. YOLOv8n), and mAP50:95 scores are 60.2% and 60.4% (+&#xa0;5.4%, +2.8%). This method achieves leading performance among 12 mainstream state-of-the-art (SOTA) detection models while maintaining high real-time efficiency, providing an effective practical solution for industrial applications in photovoltaic quality inspection systems.</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

An efficient dual-branch multi-scale approach for photovoltaic defect detection in electroluminescence images

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Yuanzhang Fan,
  • Kang Duan,
  • Yanshan Zhang

摘要

Electroluminescence (EL) imaging is currently the most efficient method for defect detection in photovoltaic modules. However, existing methods still have limitations in addressing complex industrial scenarios, failing to meet high-precision and real-time quality inspection requirements. This paper proposes a novel dual-branch multi-scale aggregation detection method based on improved YOLOv8, aiming to solve three key challenges in EL image defect detection: shape variations of homologous defects, polysilicon substrate background complexity (including overlapping defect identification), and high local detail similarity among different defect types. The framework has three core innovations: (1) its dual-branch backbone uses a parallel coarse-grained branch to capture global features of large-area defects and a fine-grained branch to focus on local details of tiny defects; their complementarity resolves feature adaptation issues caused by homologous defect shape differences; (2) the lightweight coordinate attention fusion module integrates spatial coordinate encoding into attention mechanisms to accurately locate defects in complex backgrounds and reduce polysilicon substrate background texture interference; (3) the designed cross-spatial feature pooling module enhances multi-scale feature distinguishability and captures subtle differences between similar defects via parallel multi-round max pooling and efficient multi-scale attention (EMA), improving the model’s classification accuracy for similar defects. Integrating these modules with an optimized YOLOv11 detection head, the system performs outstandingly on the PVEL-AD-2021 and PV-Multi-Defect datasets: mAP50 scores reach 90.1% and 83.8% (+ 8.7%, + 5.1% vs. YOLOv8n), and mAP50:95 scores are 60.2% and 60.4% (+ 5.4%, +2.8%). This method achieves leading performance among 12 mainstream state-of-the-art (SOTA) detection models while maintaining high real-time efficiency, providing an effective practical solution for industrial applications in photovoltaic quality inspection systems.