MSEA-YOLO: a multi-scale edge feature fusion and attention-enhanced framework for photovoltaic cell defect detection

Abstract Defects in photovoltaic cells can reduce power generation efficiency by 5%–30% and pose operational safety risks, necessitating accurate and efficient automated inspection solutions. However, existing detection methods face three key challenges: (1) insufficient localization accuracy for microscopic defects, (2) limited multi-scale feature fusion capability, and (3) inadequate small object detection robustness. This study proposes MSEA-YOLO, an improved YOLOv11-based algorithm integrating three complementary modules: C3k2-MSDR for multi-scale dilated feature extraction with structural reparameterization, multi-scale edge feature fusion for Sobel-based edge feature fusion, and attention-based intra-scale feature interaction for global contextual modeling via self-attention. On the PVEL-AD dataset, MSEA-YOLO achieves 93.4% mAP@0.5 and 63.4% mAP@0.5:0.95, improving the YOLOv11n baseline by 2.3 percentage points in mAP@0.5 and 3.4 percentage points in mAP@0.5:0.95, respectively, while maintaining 91.3% precision and 86.1% recall. Cross-dataset evaluation on the ELPV dataset confirms the architectural generalization capability of the proposed method, achieving 76.0% mAP@0.5 with a notable 6.9 percentage point improvement in precision. MSEA-YOLO outperforms YOLOv13n by 2.6 and 1.9 percentage points in mAP@0.5 on PVEL-AD and ELPV datasets, respectively. With only 3.9 M parameters and an inference speed of 377 frames per second, MSEA-YOLO achieves competitive accuracy at significantly lower computational cost, demonstrating its suitability for resource-constrained industrial deployment.