A YOLOv11 Photovoltaic Cell EL Defect Detection Method Guided by Defect Priors
Abstract
Photovoltaic cell defect detection based on electroluminescence (EL) images is essential for improving the reliability and quality control efficiency of photovoltaic module manufacturing. However, industrial EL images often contain small-scale defects, weak-texture anomalies, complex background interference, and imbalanced defect categories, which make accurate detection challenging for conventional object detection models. To address these issues, this paper proposes PGS-YOLOv11, a defect-prior-guided YOLOv11 framework for photovoltaic cell EL defect detection. Different from generic feature enhancement methods, PGS-YOLOv11 explicitly models defect-specific morphological and grayscale priors, including edge discontinuity, local contrast variation, and grayscale anomaly, and introduces them into multi-scale feature learning. Specifically, a defect prior-guided attention module is designed to transform these priors into spatial guidance, enabling the network to focus on suspected defect regions. A P2 small-object detection head is further introduced to preserve shallow high-resolution details and improve the detection of fine defects such as cracks, scratches, and finger interruptions. In addition, Focal Loss is adopted to enhance the learning of hard samples and minority defect categories. Experiments on the PVEL-AD dataset show that the proposed method achieves 0.941 Precision, 0.867 Recall, 0.933 mAP@0.5, and 0.782 mAP@0.5:0.95. Compared with YOLOv11n, PGS-YOLOv11 improves Precision, mAP@0.5, and mAP@0.5:0.95 by 0.090, 0.071, and 0.140, respectively. The results demonstrate that the proposed method improves both defect recognition accuracy and localization quality for photovoltaic cell EL defect detection.