期刊检索

  • 2026年第58卷
  • 2025年第57卷
  • 2024年第56卷
  • 2023年第55卷
  • 2022年第54卷
  • 2021年第53卷
  • 2020年第52卷
  • 2019年第51卷
  • 2018年第50卷
  • 2017年第49卷
  • 2016年第48卷
  • 2015年第47卷
  • 2014年第46卷
  • 2013年第45卷
  • 2012年第44卷
  • 2011年第43卷
  • 2010年第42卷
  • 第1期
  • 第2期

主管单位 中华人民共和国工业和信息化部 主办单位 哈尔滨工业大学 主编 李隆球 国际刊号ISSN 0367-6234 国内刊号CN 23-1235/T

期刊网站二维码
微信公众号二维码
引用本文:马保全,范文龙,杨煜晨,韩昱哲,郝博文,武晓文,王甜竹.融合混合池化与局部注意力的无人机目标检测[J].哈尔滨工业大学学报,2026,58(8):154.DOI:10.11918/202511105
Ma Baoquan,Fan Wenlong,Yang Yuchen,Han Yuzhe,Hao Bowen,Wu Xiaowen,Wang Tianzhu.Unmanned aerial vehicle object detection with hybrid pooling and local attention[J].Journal of Harbin Institute of Technology,2026,58(8):154.DOI:10.11918/202511105
【打印本页】   【HTML】   【下载PDF全文】   【查看/发表评论】  【下载PDF阅读器】  【关闭】
过刊浏览    高级检索
本文已被:浏览 946次   下载 101次 本文二维码信息
码上扫一扫!
分享到: 微信 更多
融合混合池化与局部注意力的无人机目标检测
马保全1,2,范文龙1,杨煜晨3,韩昱哲4,郝博文2,武晓文2,王甜竹2
(1.北京理工大学 计算机学院,北京 100081; 2.华北计算机系统工程研究所,北京 100000;3.南京理工大学 紫金学院,南京 210023; 4.中国民航大学 空中交通管理学院,天津 300300)
摘要:
为解决无人机航拍图像中目标尺度变化剧烈、背景复杂,以及小目标占比高导致现有YOLO系列模型单阶段检测算法小、目标检测性能不足的问题,提出一种融合混合池化与局部注意力机制的无人机目标检测方法HP-LS-YOLO。该方法以YOLOv11模型为基础框架,首先,设计空间金字塔混合池化快速策略(SPHPF),在单一尺度上并行融合最大池化与平均池化特征,引入多感受野上下文信息,并通过卷积自适应筛选增强复杂背景下的特征表达能力。其次,构建融合小波卷积与局部注意力机制的C2PLSA模块,其中小波变换注意力在保持模型轻量化的前提下有效扩大感受野,捕获小目标周围的大尺度场景语义;局部自注意力块采用Top-K稀疏注意力机制,突出小目标的局部边缘与纹理特征,同时抑制冗余背景干扰。最后,在多个公开数据集上开展对比验证实验,通过与原模型以及同类型轻量化检测算法的对比,全方位检验所提改进策略对无人机航拍密集小目标检测效果的提升作用。结果表明:与YOLOv11模型相比,所提方法在MSCOCO数据集验证集上的APval50∶95、APval50、APvalS分别提升 0.9%、0.9%、1.4%,在VisDrone和UAVDT数据集上的额外实验进一步验证了该方法在复杂无人机场景下的鲁棒性,所提出的HP-LS-YOLO模型能够在保持模型轻量化的同时,有效提升无人机航拍图像中密集小目标的检测性能。
关键词:  YOLOv11  无人机航拍图像  小目标检测  混合池化  局部注意力机制
DOI:10.11918/202511105
分类号:TP391.41
基金项目:中国教育部基础学科与交叉学科突破计划(JYB2025XDXM108、JYB2025XDXM201);国家自然科学基金(62225203、62532001);北京市自然科学基金(L241010)
Unmanned aerial vehicle object detection with hybrid pooling and local attention
Ma Baoquan1,2,Fan Wenlong1,Yang Yuchen3,Han Yuzhe4,Hao Bowen2,Wu Xiaowen2,Wang Tianzhu2
(1.School of Computer Science & Technology, Beijing Institute of Technology, Beijing 100081, China; 2.North China Computer System Engineering Research Institute, Beijing 100000, China; 3.Zijin College, Nanjing University of Science and Technology, Nanjing 210023, China; 4.College of Air Traffic Management, Civil Aviation University of China, Tianjin 300300, China)
Abstract:
To solve the problem of insufficient small object detection performance of existing YOLO series single-stage detection algorithms caused by severe object scale variation, complex backgrounds, and a high proportion of small objects in unmanned aerial vehicle aerial images, this paper proposed an unmanned aerial vehicle object detection method fusing hybrid pooling and local attention mechanism, named HP-LS-YOLO. Based on the YOLOv11 framework, this method first designed a spatial pyramid hybrid pooling fast (SPHPF) strategy to introduce multi-receptive field contextual information at a single scale by parallelly fusing max pooling and average pooling features and enhanced the feature representation capability under complex backgrounds through convolution-based adaptive filtering. Secondly, a C2PLSA module integrating wavelet convolution and local attention mechanism was constructed, in which the wavelet transform attention effectively enlarged the receptive field while keeping the model lightweight and introduced large-scale scene semantics around small objects. The local self-attention block adopted a Top-K sparse attention mechanism to highlight the local edge and texture features of small objects, while suppressing redundant background interference. Finally, comparative validation experiments were carried out on multiple public datasets to comprehensively verify the enhancement effect of the proposed improvement strategies on dense small object detection in unmanned aerial vehicle aerial images by comparing with the baseline model and similar lightweight detection algorithms. The results indicate that compared with YOLOv11, the proposed method improves APval50∶95, APval50, APvalS on the validation set of the MSCOCO dataset by 0.9%, 0.9%, 1.4%, respectively. Additional experiments on VisDrone and UAVDT datasets further verify the robustness of this method in complex unmanned aerial vehicle scenarios. The proposed HP-LS-YOLO can effectively improve the detection performance of dense small objects in unmanned aerial vehicle aerial images while keeping the model lightweight.
Key words:  YOLOv11  unmanned aerial vehicle aerial image  small object detection  hybrid pooling  local attention mechanism

友情链接LINKS