YOLOv9s-PILD: A Lightweight Photovoltaic Module Defect Detection Method Based on Grid-Structure Priors

Detecting fine cracks and hot spots in photovoltaic modules is challenging due to their weak and small-scale features, which are easily obscured by regular background textures such as grid lines and cell boundaries. Consequently, general-purpose detection networks struggle to distinguish these defects, and their large model sizes hinder deployment on edge devices. To address these issues, this paper proposes YOLOv9s-PILD, a lightweight detection model built upon the YOLOv9s baseline. The proposed model incorporates three key innovations: a LiteNeck architecture to reduce redundancy in feature fusion; a Photovoltaic Grid Anomaly Enhancement (PVGA) that leverages prior knowledge of module structures to enhance defect features and suppress interference; and a Location-Aware Weighting Module (LAW) combined with a Lightweight Decoupled Detection Head (LDHead) to improve localization accuracy and reduce task coupling. Experimental results demonstrate that the model achieves 5.27M parameters and 19.30 GFLOPs, attaining an mAP@0.5 of 81.18%—a 1.81 percentage point improvement over YOLOv9s. These findings indicate that the proposed method improves detection performance while reducing model complexity.