Defect Detection in PV Electroluminescence Images Using YOLOv9 to YOLOv12 Lightweights Variants

Accurate defect detection and classification are essential for ensuring the optimal performance and efficiency of photovoltaic (PV) panels. In this study, we evaluate the performance of the latest lightweight YOLO architectures, specifically variants 9 to 12, including their tiny/nano and small variants, on the PVEL-AD electroluminescence dataset. This dataset comprises 12 distinct defect classes, and we report both overall and per-class performance metrics. To address the inherent class imbalance in the PVEL-AD dataset, which can compromise the reliability of results, we implemented a dual augmentation strategy combining YOLO augmentation with CycleGAN-based generative data augmentation. Our experimental results demonstrate that YOLO 11s outperforms other variants, achieving a mAP@50 of 84.12% and a mAP@50-95 of 56.22%. These metrics highlight its ability to generate bounding boxes that closely align with ground truth annotations. Additionally, YOLO 11s achieved the highest F1 score of 80.19 %, indicating an optimal balance between precision and recall, while maintaining a rapid inference time of 3.3 ms.