<p>Unmanned aerial vehicle (UAV) object detection is a critical task in visual computing, underpinning applications in traffic surveillance, disaster response, and precision agriculture. However, this task is uniquely challenging due to the aerial perspective, which often renders small objects as clusters of only tens of pixels, with blurred edges and weak signals easily overwhelmed by complex, cluttered backgrounds. Current detection frameworks frequently struggle to reconcile the need for fine-grained detail capture with the stringent real-time demands of UAV platforms. To address this, we introduce SF-DETR, an end-to-end Detection Transformer framework designed for robust and efficient small-object detection in UAV imagery. Methodologically, our contribution is twofold. First, we propose a lightweight backbone network, SFH-MFF, which embeds a learnable frequency filtering module within shallow layers to explicitly amplify subtle high-frequency textures, achieving a complementary feature representation across spatial and frequency domains. Second, we design a dual-path linear attention module, DL-AFIM, within the encoder. This module employs a positive–negative perception mechanism to effectively distinguish foreground targets from background clutter, while an integrated frequency-domain feed-forward network and an adaptive multi-scale residual fusion strategy enhance high-frequency detail representation without increasing parameter burden. Here we show that on the VisDrone and UAVDT datasets, our method achieves a 1.8% and 1.5% improvement in AP, and a 2.6% and 2.1% gain in <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {AP}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>AP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation>, respectively, compared to the RT-DETR baseline. Furthermore, the model operates at up to 112 FPS on an RTX 4090 GPU, satisfying real-time constraints. Experiments on the SIMD remote sensing dataset further validate the strong cross-dataset generalization capability of our approach. This work offers a new pathway for integrating frequency-domain analysis within efficient transformer architectures, providing a robust solution for real-time aerial perception. Our code is publicly available at: <a href="https://github.com/twy-ui/SF-DETR">https://github.com/twy-ui/SF-DETR</a>( DOI:10.5281/zenodo.18883197 ).</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

Enhancing UAV small-object detection via spatial–frequency synergy and polarity-aware attention

  • Weiyan Tang,
  • Fuzhen Sun,
  • Zihao Jing,
  • Zhuangrui Zhu,
  • Yudong Li,
  • Shaoqing Wang

摘要

Unmanned aerial vehicle (UAV) object detection is a critical task in visual computing, underpinning applications in traffic surveillance, disaster response, and precision agriculture. However, this task is uniquely challenging due to the aerial perspective, which often renders small objects as clusters of only tens of pixels, with blurred edges and weak signals easily overwhelmed by complex, cluttered backgrounds. Current detection frameworks frequently struggle to reconcile the need for fine-grained detail capture with the stringent real-time demands of UAV platforms. To address this, we introduce SF-DETR, an end-to-end Detection Transformer framework designed for robust and efficient small-object detection in UAV imagery. Methodologically, our contribution is twofold. First, we propose a lightweight backbone network, SFH-MFF, which embeds a learnable frequency filtering module within shallow layers to explicitly amplify subtle high-frequency textures, achieving a complementary feature representation across spatial and frequency domains. Second, we design a dual-path linear attention module, DL-AFIM, within the encoder. This module employs a positive–negative perception mechanism to effectively distinguish foreground targets from background clutter, while an integrated frequency-domain feed-forward network and an adaptive multi-scale residual fusion strategy enhance high-frequency detail representation without increasing parameter burden. Here we show that on the VisDrone and UAVDT datasets, our method achieves a 1.8% and 1.5% improvement in AP, and a 2.6% and 2.1% gain in \(\hbox {AP}_{50}\) AP 50 , respectively, compared to the RT-DETR baseline. Furthermore, the model operates at up to 112 FPS on an RTX 4090 GPU, satisfying real-time constraints. Experiments on the SIMD remote sensing dataset further validate the strong cross-dataset generalization capability of our approach. This work offers a new pathway for integrating frequency-domain analysis within efficient transformer architectures, providing a robust solution for real-time aerial perception. Our code is publicly available at: https://github.com/twy-ui/SF-DETR( DOI:10.5281/zenodo.18883197 ).