Abstract
<title>Abstract</title> <p>With the rapid expansion of photovoltaic (PV) power generation, the demand for defect detection in PV modules continues to evolve. However, most existing PV inspection models rely on images acquired through specialized imaging technologies such as electroluminescence (EL) and infrared (IR) imaging, which require costly equipment and are difficult to scale. There remains a lack of cost-effective solutions suitable for large-scale deployment. In this work, we focus on visible light imaging as a practical alternative and develop a detection model that achieves a favorable balance between accuracy and lightweight design. We propose a novel detection framework, CCM-YOLO, which integrates multiple structural innovations to better meet the requirements of PV defect detection. Specifically, a new spatial-channel separated downsampling method is introduced to strengthen feature extraction while reducing computational load. To further improve model performance and reduce size, two efficient feature extraction modules based on the Cross Stage Partial (CSP) architecture are designed. The first module, CSP with Mixed Local Channel Attention (CSPMLCA), incorporates Mixed Local Channel Attention (MLCA) into the Neck to improve feature representation without increasing model size or parameter count. The second module, the CSP with HetConv (CSPHet), replaces the original Backbone with a structure built using HetConv, effectively reducing model parameters. Experimental results on public photovoltaic defect datasets demonstrate that the proposed model outperforms the baseline and several state-of-the-art object detection methods across multiple evaluation metrics. The CCM-YOLO model achieves a mean average precision (mAP) of 95.19% and a detection precision of 93.19%. Compared to the baseline model, CCM-YOLO improves mAP by 0.73%, detection precision by 0.3%, reduces model size by 21.8%, and cuts parameter count by 22.5%, with a final model size of only 17.6 MB. CCM-YOLO achieves an effective trade-off between detection accuracy and model size, demonstrating strong lightweight characteristics suitable for deployment in resource-constrained environments.</p>