Research on surface defect detection of PV modules based on lightweight YOLOv10s

To address the issues of low detection accuracy and large model size in surface defect detection of PV modules, this paper proposes a lightweight improved YOLOv10s model. First, the backbone network is replaced with StarNet, leveraging its advantage of computing high-dimensional features in low-dimensional spaces to enhance feature extraction capability for minor defects. Second, the C2f-Faster-EMA module is introduced to replace the original C2f module in the neck, reducing redundant computations and enabling efficient feature aggregation. Finally, the Detect-LSCSBD detection head is designed, incorporating a shared convolution mechanism and independent batch normalization to reduce parameters and improve global information fusion between feature maps. Experimental results show that when the improved YOLOv10s model is trained on a public dataset of PV module surface defects, it achieves detection of broken, durty, and clean PV modules. The model’s parameter count and computational load are reduced by 45.4% and 41.1%, respectively, while the mAP50 is increased by 2.2%. The improved model is suitable for deployment on edge devices such as drones for defect detection in PV modules.