Abstract: In order to address the issues of low accuracy, large model parameters, and the occurrences of missed detections and false detections under complex backgrounds in the existing traditional photovoltaic panel defect detection, this paper proposes an efficient photovoltaic panel defect detection algorithm based on the RT-DETR model. Firstly, to boost detection accuracy, the FREBlock architecture is developed, which not only improves feature extraction but also enhances detection efficiency. Secondly, the CRDFP multi-scale feature fusion structure is designed to strengthen the integration of features across different scales. Lastly, the deformable attention mechanism is incorporated, which enables the model to focus on the information features of the region of interest. The experimental results indicate that the improved model achieves an average mean Average Precision (mAP) of 79.2%, an increase of 3.6 percentage points over the traditional model. Additionally, the model's parameters reduces 22.6%, and the computational load decreases 25.9%, demonstrating a high capacity for real-time detection.

Key words: deep learning, RT-DETR, photovoltaic panels, defect detection, multi-scale feature fusion