基于改进RT-DETR的空海无人机目标检测模型研究

Research on Air-Sea UAV Target Detection Model Based on Improved RT-DETR

摘要: 针对空海无人机目标检测中目标尺度变化大、动态背景干扰强，以及传统模型泛化能力不足、多尺度特征处理效率低的问题，本文提出一种融合频域解耦与视觉Transformer的RT-DETR改进模型(FD-ViT-RT-DETR)。该模型以RT-DETR为基础框架，前端设计频域解耦模块，通过快速傅里叶变换分离图像域不变特征(目标核心特征)与域特定特征(环境干扰特征)，并结合实例级对比损失强化特征区分度；中间优化视觉Transformer架构，新增空间压缩网络降低30%计算成本以适配边缘设备，同时引入位置注意力偏置提升长距离依赖捕捉能力；后端搭配不确定性最小查询选择方案优化检测精度。在含508张空海无人机图像的数据集上实验，结果显示：该模型精度(P)达86.4%、召回率(R)达87.4%、mAP50达88.1%，均优于YOLOv5、YOLOv8及原始RT-DETR，且在强逆光、海浪反光等复杂场景中鲁棒性更强，可为空海搜救、安防监控等实际应用提供高效检测方案。

Abstract: Aiming at the problems in air-sea UAV target detection, such as large target scale variation, strong dynamic background interference, insufficient generalization ability of traditional models, and low efficiency in multi-scale feature processing, this paper proposes an improved RT-DETR model integrating frequency domain decoupling and Vision Transformer (FD-ViT-RT-DETR). Based on the RT-DETR framework, the model is designed with a frequency domain decoupling module at the front end, which separates image domain-invariant features (core target features) from domain-specific features (environmental interference features) through fast Fourier transform, and enhances feature distinguishability by combining instance-level contrastive loss. In the middle layer, the Vision Transformer architecture is optimized by adding a spatial compression network to reduce 30% of computational costs for adapting to edge devices, while introducing positional attention bias to improve the ability of capturing long-range dependencies. At the back end, an uncertainty-minimized query selection scheme is adopted to optimize detection accuracy. Experiments on a dataset containing 508 air-sea UAV images show that the model achieves a precision (P) of 86.4%, a recall (R) of 87.4%, and an mAP50 of 88.1%, all outperforming YOLOv5, YOLOv8, and the original RT-DETR. It also exhibits stronger robustness in complex scenarios such as strong backlight and sea wave reflection, providing an efficient detection solution for practical applications like air-sea search and rescue, and security monitoring.