<p>Electroluminescence (EL) imaging is a core tool for quality inspection in photovoltaic (PV) manufacturing, where micro-cracks, finger interruptions, black-core defects, and dislocations can severely degrade module efficiency and reliability. Automatic EL defect detection remains challenging because defects appear under extremely low contrast, are corrupted by strong sensor noise, exhibit highly irregular and anisotropic geometry, and often differ only by subtle texture and shape cues. To address these difficulties, we propose YOLO-PVELAD, a task-specialized one-stage detector tailored to EL imagery and the PVELAD EL2021 benchmark. The method exploits the fact that EL defects are dominated by high-frequency, crack-like patterns on globally smooth backgrounds and that semantic categories and geometric support are only weakly correlated. YOLO-PVELAD integrates three tightly coupled components: a Directional Crack-aware Attention (DCA) module that uses directional pooling and one-dimensional convolutions along horizontal and vertical axes to efficiently enhance anisotropic defect structures; a DCT High-Frequency Feature Pyramid Network (DHF-FPN) that injects Discrete Cosine Transform (DCT)-based high-frequency cues into multi-scale features to improve the visibility and separability of subtle anomalies; and a Geometry-Aligned Dual-branch Head (GA-Head) that decouples classification and regression, employs dynamic deformable sampling to adapt receptive fields to irregular defect shapes, and performs geometry-aware prediction. Experiments on PVELAD EL2021 show that YOLO-PVELAD achieves 76.3% mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> and 48.6% mAP<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$_{50{:}95}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow> <mn>50</mn> <mo>:</mo> <mn>95</mn> </mrow> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> in a representative single-run setting (seed=42), outperforming a strong YOLO11 baseline by +14.8 mAP<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> and +18.0 recall, while maintaining real-time throughput. These results confirm the effectiveness of directional crack-aware attention, explicit high-frequency modeling, and geometry-aligned dual-branch regression for robust EL defect detection, and indicate their potential for broader high-frequency-dominant industrial vision tasks.</p>

Geometry-aligned dual-branch detection with directional crack-aware attention for EL imaging of photovoltaic cells

期刊论文
出版模式：

开放获取
发表日期：

2026年2月28日

查看全文

Jincheng Li,
Fugui Sun,
Liang Cheng

摘要

Electroluminescence (EL) imaging is a core tool for quality inspection in photovoltaic (PV) manufacturing, where micro-cracks, finger interruptions, black-core defects, and dislocations can severely degrade module efficiency and reliability. Automatic EL defect detection remains challenging because defects appear under extremely low contrast, are corrupted by strong sensor noise, exhibit highly irregular and anisotropic geometry, and often differ only by subtle texture and shape cues. To address these difficulties, we propose YOLO-PVELAD, a task-specialized one-stage detector tailored to EL imagery and the PVELAD EL2021 benchmark. The method exploits the fact that EL defects are dominated by high-frequency, crack-like patterns on globally smooth backgrounds and that semantic categories and geometric support are only weakly correlated. YOLO-PVELAD integrates three tightly coupled components: a Directional Crack-aware Attention (DCA) module that uses directional pooling and one-dimensional convolutions along horizontal and vertical axes to efficiently enhance anisotropic defect structures; a DCT High-Frequency Feature Pyramid Network (DHF-FPN) that injects Discrete Cosine Transform (DCT)-based high-frequency cues into multi-scale features to improve the visibility and separability of subtle anomalies; and a Geometry-Aligned Dual-branch Head (GA-Head) that decouples classification and regression, employs dynamic deformable sampling to adapt receptive fields to irregular defect shapes, and performs geometry-aware prediction. Experiments on PVELAD EL2021 show that YOLO-PVELAD achieves 76.3% mAP $_{50}$ $_{50}$ and 48.6% mAP $_{50{:}95}$ $_{50 : 95}$ in a representative single-run setting (seed=42), outperforming a strong YOLO11 baseline by +14.8 mAP $_{50}$ $_{50}$ and +18.0 recall, while maintaining real-time throughput. These results confirm the effectiveness of directional crack-aware attention, explicit high-frequency modeling, and geometry-aligned dual-branch regression for robust EL defect detection, and indicate their potential for broader high-frequency-dominant industrial vision tasks.