Research on UAV Recognition Method Based on DETR Algorithm

In recent years, the rapid development of UAV technology has led to its widespread application in agriculture, military operations, and surveillance. However, increasing UAV use poses serious security threats such as smuggling and unauthorized reconnaissance. A key challenge in UAV recognition lies in recognizing small targets in complex environments, where traditional methods often lack accuracy and adaptability. Deep learning-based approaches offer improvements but still face difficulties in handling scale variation and small object recognition. This paper presents an improved recognition model based on DETR and RT-DETR architectures. The model integrates multi-scale feature fusion, a lightweight structure, and an optimized backbone enhanced by an Exponential Moving Average (EMA) module to improve recognition performance. Experiments were conducted on the Det-Fly dataset, where our improved RT-DETR achieved a precision of 97.3% (a 0.4% increase over the original RT-DETR), a recall of 88.2% (a 0.5% increase), an mAP@50 of 92.2%, and an mAP@50–95 of 56.1%. These results confirm the method’s effectiveness in terms of accuracy, robustness, and real-time performance, showing strong potential for real-world UAV recognition applications.