With the rapid development of photovoltaic (PV) power generation technology, defect detection in PV modules is crucial for the operational efficiency and energy security of power plants. However, traditional detection methods are often limited when processing aerial images, as they are prone to issues such as multi-scale feature blurring and small target missed detection, making it difficult to meet the high precision and real-time engineering requirements. To address these issues, an optimized model based on YOLOv8(you only look once v8), named RD (recursive depthwise)-YOLO, was proposed. A multi-scale channel squeeze-and-excitation network (MultiSENet) was introduced to enhance its ability to capture multi-scale features. Additionally, a recursive depthwise separable convolutional block attention module (RDCBAM) was incorporated to optimize channel information and improve the model's detection accuracy for small targets. Experiments were conducted using a self-constructed dataset, and the results indicate that RD-YOLO significantly improves both the accuracy and efficiency compared to traditional detection models.