辽宁石油化工大学学报 ›› 2026, Vol. 46 ›› Issue (4): 79-87.DOI: 10.12422/j.issn.1672-6952.2026.04.010

• 信息与控制工程 • 上一篇    下一篇

基于YOLO11⁃RHO的无人机航拍图像小目标检测算法

岳思洋1(), 佐安帆1(), 石元博1, 黄越洋2, 刘思妍1   

  1. 1.辽宁石油化工大学 人工智能与软件学院,辽宁 抚顺 113001
    2.辽宁石油化工大学 信息与控制工程学院,辽宁 抚顺 113001
  • 收稿日期:2026-03-01 修回日期:2026-03-30 出版日期:2026-08-25 发布日期:2026-07-20
  • 通讯作者: 佐安帆
  • 作者简介:岳思洋(2000-),女,硕士研究生,从事目标检测方面的研究;E⁃mail:1577294702@qq.com
  • 基金资助:
    辽宁省教育厅创新发展基金项目(LJ242510148001)

YOLO11⁃RHO: A Small⁃Object Detection Algorithm for UAV Aerial Images

Siyang YUE1(), Anfan ZUO1(), Yuanbo SHI1, Yueyang HUANG2, Siyan LIU1   

  1. 1.School of Artificial Intelligence and Software,Liaoning Petrochemical University,Fushun Liaoning 113001,China
    2.School of Information and Control Engineering,Liaoning Petrochemical University,Fushun Liaoning 113001,China
  • Received:2026-03-01 Revised:2026-03-30 Published:2026-08-25 Online:2026-07-20
  • Contact: Anfan ZUO

摘要:

无人机航拍图像中小目标存在像素占比低、易受复杂背景干扰且尺度变化大等问题,导致现有检测易出现漏检与精度不足的情况。为此,在YOLOv11算法的基础上提出了面向无人机航拍场景的YOLO11⁃RHO改进算法。首先,在骨干网络中引入带有频率先验初始化的重参数化卷积(Online Convolutional Re⁃parameterization,OREPA),利用空间频率选择性增强模型对高频边缘和纹理信息的表征能力;其次,采用RepNCSPELAN4_high模块优化梯度传递路径,构建梯度直达通路以减轻小目标特征在深层网络中的衰减;再次,在特征融合网络中集成HFFB分层特征融合块,结合空间与通道解耦的自适应注意力机制,强化小目标表征,实现多尺度信息对齐;最后,面向轻量与实时部署进行整体设计。在VisDrone2019数据集上的对比实验与消融实验结果表明,改进模型性能优异,mAP@0.5=0.401,mAP@0.50∶0.95=0.245;同时,保持了轻量化架构(参数量为6.22 M,计算量为15.1 GFLOPs),有效兼顾了高检测精度与边缘端实时运行的需求,满足了端侧实际部署的要求。

关键词: 无人机, 小目标检测, YOLOv11, 重参数化卷积, RepNCSPELAN4, 特征融合

Abstract:

Small objects in UAV aerial images have low pixel proportions, are vulnerable to interference from complex backgrounds, and exhibit large scale variations, which often lead to missed detections and insufficient accuracy in existing methods. Therefore, an improved algorithm, YOLO11⁃RHO, is proposed for UAV aerial imagery based on YOLOv11. First, online convolutional re⁃parameterization (OREPA) with frequency⁃prior initialization is introduced into the backbone network to enhance the model's representation of high⁃frequency edge and texture information through spatial⁃frequency selectivity. Second, the RepNCSPELAN4_high module is used to optimize gradient propagation paths and construct a direct gradient pathway, thereby alleviating the attenuation of small⁃object features in deep network layers. Third, a hierarchical feature fusion block (HFFB) is integrated into the feature⁃fusion network. By combining a spatial⁃ and channel⁃decoupled adaptive attention mechanism, the block strengthens small⁃object representation and achieves multiscale information alignment. Finally, the overall architecture is designed for lightweight real⁃time deployment. Comparative and ablation experiments on the VisDrone2019 dataset show that the improved model achieves excellent performance, with mAP@0.5=0.401 and mAP@0.50∶0.95=0.245. It also maintains a lightweight architecture with 6.22 M parameters and 15.1 GFLOPs, effectively balancing high detection accuracy with real⁃time edge⁃device operation and satisfying practical end⁃side deployment requirements.

Key words: UAV, Small-object detection, YOLOv11, Re-parameterized convolution, RepNCSPELAN4, Feature fusion

中图分类号: 

引用本文

岳思洋, 佐安帆, 石元博, 黄越洋, 刘思妍. 基于YOLO11⁃RHO的无人机航拍图像小目标检测算法[J]. 辽宁石油化工大学学报, 2026, 46(4): 79-87.

Siyang YUE, Anfan ZUO, Yuanbo SHI, Yueyang HUANG, Siyan LIU. YOLO11⁃RHO: A Small⁃Object Detection Algorithm for UAV Aerial Images[J]. Journal of Liaoning Petrochemical University, 2026, 46(4): 79-87.

使用本文

0
    /   /   推荐

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://journal.lnpu.edu.cn/CN/10.12422/j.issn.1672-6952.2026.04.010

               https://journal.lnpu.edu.cn/CN/Y2026/V46/I4/79