YOLO-LitePV: a lightweight detection algorithm for photovoltaic panel defects

Defects on photovoltaic (PV) cells can severely compromise the power generation efficiency and service life of photovoltaic modules. To address the low operational efficiency of detection algorithms and the low accuracy due to the similarity and large-scale variance of PV defects, we propose an improved lightweight detection model, YOLO-LitePV, based on the YOLOv8 architecture. The model enhances the feature recognition capability of multi-scale defects by incorporating the Simple, Parameter-Free Attention Module (SimAM) and Efficient Multi-Scale Attention (EMA) module. It also achieves model lightweighting through structured channel pruning based on the dependency graph. Additionally, knowledge distillation technology based on a trust gate mechanism is employed to compensate for the minor accuracy loss caused by pruning, thereby retaining the model’s high detection accuracy. Experimental results on the PVEL-AD dataset demonstrate that YOLO-LitePV achieves an mAP50 of 90.3% and an FPS of 33.65 when deployed on an Intel N100 embedded PC. Compared to the original YOLOv8n model, there is a 2.8% increase in mAP50 and the FPS is improved to 1.83 times the original, providing an efficient solution for industrial defect detection in photovoltaic panels.