Abstract: To address the issues of weakened features for distant small-target defects and high model complexity in photovoltaic panel defect detection algorithms, this study proposes an improved lightweight algorithm named FEM-YOLO. Firstly, the C3k2 module is enhanced by integrating FasterBlock and EMA, constructing a C3k2-Faster-EMA structure to improve the network's ability to learn and capture features of defective targets. Subsequently, the Mona module is incorporated into the C2PSA block, optimizing the model's feature extraction and representation capabilities. Moreover, the MLCA mechanism is integrated into the backbone network to enhance the robustness of feature extraction for diverse targets. Finally, an additional P2 detection layer is added specifically for small targets, and an efficient detection head named EfficientHead is designed. This combination enhances the capability to capture micro-defects while simultaneously reducing model complexity. Experimental results demonstrate that, compared with the original YOLO11n model, the improved algorithm achieves increases of 1.9% in both mAP50 and mAP50-95 metrics. Furthermore, the model parameter count is reduced to 2.1×10⁶ and the model size is compressed to 4.4 MiB. Thus, the proposed FEM-YOLO algorithm significantly enhances detection accuracy while substantially reducing model complexity.

Key words: photovoltaic panel, defect detection, lightweight model, small object detection, YOLO11n