Abstract: Surface defect detection in photovoltaic (PV) cells is crucial for ensuring the safe and efficient operation of solar power systems. This paper proposes an enhanced RT-DETR model, named CAD-RTDETR, for efficient defect detection on PV cell surfaces. First, the re-parameterized module ConvCX3, derived from a parameter-free attention mechanism network, is introduced to improve detection accuracy. Second, the improved ASF-YOLO object detection model is integrated to optimize multi-scale feature extraction. Third, the DySample dynamic upsampler is incorporated to enhance anti-interference capabilities and refine detection performance. Finally, the Inner-ShapeIoU loss function is adopted to improve detection accuracy for small targets with complex shapes and varying scales. Experimental results demonstrate that CAD-RTDETR outperforms the baseline RT-DETR-r18 in detection precision, speed and robustness. Specifically, it achieves a 3.70% increase in accuracy, a 5.10% improvement in recall rate, and a 7.00% boost in mean average precision (mAP) for small defect detection. Comparative trials on the PVEL-AD dataset confirm the model’s superior generalization capability over conventional algorithms. These advancements provide an efficient and accurate solution for PV cell surface defect detection, offering significant practical value for industrial applications.