Scholarlink

<p>Object detection in unmanned aerial vehicle (UAV) imagery is challenging due to extremely tiny object sizes, complex backgrounds, and large scale variations. To address these challenges, we propose an end-to-end spatial–frequency dual-guided multi-scale network for tiny object detection in UAV aerial images (SFD-DETR). First, we propose a lightweight Multi-Scale Channel Separation (MSCS) feature extraction network to enhance representation efficiency while reducing model complexity. Furthermore, we propose a Spatial–Frequency Enhancement Module (SFEM) to preserve fine-grained spatial details and improve tiny object discriminability through adaptive global–local feature recalibration. In addition, we employ a frequency-adaptive upsampling operator, Converse2D (C2D), together with the Shape-IoU loss to mitigate high-frequency detail loss and improve bounding box regression. Experiments on the VisDrone and HIT-UAV benchmarks demonstrate that SFD-DETR achieves 39.4% and 83.1% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mn>50</mn> </msub> </mrow> </math></EquationSource> </InlineEquation>, respectively, outperforming RT-DETR by 3.2 and 3.6 percentage points while reducing parameters by 20.2%. Furthermore, the high computational density of our frequency-domain operations intrinsically leverages parallel processing paradigms, making it well suited for high-throughput UAV data streams on modern High-Performance Computing (HPC) architectures. Our code will be available at https://github.com/arvinooo/SFD-DETR-main</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SFD-DETR: an end-to-end spatial–frequency dual-guided multi-scale network for tiny object detection in UAV aerial images

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Weiquan Li,
  • Zhijie Xu,
  • Jianqin Zhang,
  • Yiting Li

摘要

Object detection in unmanned aerial vehicle (UAV) imagery is challenging due to extremely tiny object sizes, complex backgrounds, and large scale variations. To address these challenges, we propose an end-to-end spatial–frequency dual-guided multi-scale network for tiny object detection in UAV aerial images (SFD-DETR). First, we propose a lightweight Multi-Scale Channel Separation (MSCS) feature extraction network to enhance representation efficiency while reducing model complexity. Furthermore, we propose a Spatial–Frequency Enhancement Module (SFEM) to preserve fine-grained spatial details and improve tiny object discriminability through adaptive global–local feature recalibration. In addition, we employ a frequency-adaptive upsampling operator, Converse2D (C2D), together with the Shape-IoU loss to mitigate high-frequency detail loss and improve bounding box regression. Experiments on the VisDrone and HIT-UAV benchmarks demonstrate that SFD-DETR achieves 39.4% and 83.1% \(mAP_{50}\) m A P 50 , respectively, outperforming RT-DETR by 3.2 and 3.6 percentage points while reducing parameters by 20.2%. Furthermore, the high computational density of our frequency-domain operations intrinsically leverages parallel processing paradigms, making it well suited for high-throughput UAV data streams on modern High-Performance Computing (HPC) architectures. Our code will be available at https://github.com/arvinooo/SFD-DETR-main