LDV-YOLO: A Lightweight Improved YOLOv11 for Photovoltaic Cell Defect Detection

Abstract

Currently, the detection of defects in solar cells faces several challenges, including the difficulty in identifying small-sized defects, high sensitivity to complex backgrounds, and the large size of existing models. To overcome these limitations, this study proposes Laser Doppler Velocimetry-You Only Look Once (LDV-YOLO), a lightweight model for efficient defect detection. Initially, the backbone feature extraction network of YOLOv11n was reconstructed using a lightweight LCNet-G design to minimize both the model's parameters and computational cost. Furthermore, we replaced YOLOv11n's original C3k2 neck module with a VoVGSCSP module, which reduces computation while improving detection accuracy. Ultimately, we substituted YOLOv11n's original Complete-IoU (CIoU) loss with the Weighted Intersection over Union (WIoU) variant, improving the localization precision of bounding boxes. Quantitative results obtained from the PV Evolution Labs (PVEL)-AD-2021 dataset indicate that LDV-YOLO achieves 1,783,282 parameters, 4.7 Giga Floating Point Operations Per second (GFLOPs), and 84.1% mAP. Compared to YOLOv11n, the proposed model reduces the parameter count by 31%, decreases GFLOPs by 25.40%, and improves mean average precision by 3.7%. The improved algorithm exhibits superior performance in PV cell defect detection while maintaining a lightweight architecture suitable for deployment on embedded devices.