Scholarlink

<p>Target detection models have achieved remarkable progress in the defect detection of photovoltaic (PV) modules’ electroluminescence (EL) images. However, numerous complex and uncertain factors in practical detection scenarios, such as scale variation, feature similarity, and occlusion effects, pose challenges for achieving accurate and real-time detection in industrial production. To address these limitations, Dual Extraction Shared Network (DESNet), a novel target detection network based on the convolutional neural network is proposed. To reduce computational resource demands, we design Projector-Smart-PDConv (ProjSPD) and High-level 2D Feature Interaction (Hi2D-FI) in Backbone. To tackle issues of target occlusion and class imbalance, Context Guided module (CG) in Neck is designed for global information fusion. To mitigate the impact of tiny target similarity, we design Shared Group Normalization Convolutional (SGNC) head. Model evaluation is conducted on two public datasets. DESNet achieves 13.4% and 1.0% higher <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {mAP}_{0.5}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mrow> <mn>0.5</mn> </mrow> </msub> </math></EquationSource> </InlineEquation> over the CNN-based baseline within the permitted detection time, validating its practical effectiveness and strong potential for industrial deployment.</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 novel real-time image defect detection network for photovoltaic modules

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Ran Song,
  • Sheng Ding,
  • Qian Zhang,
  • Jinya Su,
  • Congyan Chen

摘要

Target detection models have achieved remarkable progress in the defect detection of photovoltaic (PV) modules’ electroluminescence (EL) images. However, numerous complex and uncertain factors in practical detection scenarios, such as scale variation, feature similarity, and occlusion effects, pose challenges for achieving accurate and real-time detection in industrial production. To address these limitations, Dual Extraction Shared Network (DESNet), a novel target detection network based on the convolutional neural network is proposed. To reduce computational resource demands, we design Projector-Smart-PDConv (ProjSPD) and High-level 2D Feature Interaction (Hi2D-FI) in Backbone. To tackle issues of target occlusion and class imbalance, Context Guided module (CG) in Neck is designed for global information fusion. To mitigate the impact of tiny target similarity, we design Shared Group Normalization Convolutional (SGNC) head. Model evaluation is conducted on two public datasets. DESNet achieves 13.4% and 1.0% higher \(\text {mAP}_{0.5}\) mAP 0.5 over the CNN-based baseline within the permitted detection time, validating its practical effectiveness and strong potential for industrial deployment.