Scholarlink

<p>Real-time small object detection in unmanned aerial vehicle imagery serves critical applications but faces significant deployment challenges. The primary constraint lies in achieving optimal trade-offs between detection accuracy and computational efficiency on resource-limited edge-AI devices. To address small object detection accuracy while maintaining computational efficiency, we propose <b>C</b>onvolutional <b>A</b>dditive <b>C</b>ontext <b>DE</b>tection <b>TR</b>ansformer (CAC-DETR), a novel detection framework based on Real-Time DEtection TRansformer (RT-DETR). Our framework introduces the CAC block, which comprises three core components: C2f, Convolutional Additive Token Mixer (CATM), and Convolutional Gated Linear Unit (CGLU). The diverse kernel sizes employed in CATM and CGLU provide flexible receptive field adjustment, while the pure convolutional architecture ensures optimal computational efficiency for edge deployment. Our proposed CAC-DETR achieves substantial performance improvements while reducing computational overhead compared to the benchmark RT-DETR. Experimental results on VisDrone demonstrate computational reductions of 30.3% in parameter complexity and 20.4% in GFLOPs, accompanied by significant accuracy improvements of 3.6% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {mAP}_{0.5}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mrow> <mn>0.5</mn> </mrow> </msub> </math></EquationSource> </InlineEquation> and 4.0% <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\text {mAP}_{0.5:0.95}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mrow> <mn>0.5</mn> <mo>:</mo> <mn>0.95</mn> </mrow> </msub> </math></EquationSource> </InlineEquation>. Consistent performance gains are also observed on HIT-UAV and DIOR datasets. Furthermore, we employ Localization Knowledge Distillation to enhance the detection accuracy of lightweight models, such as YOLOv8-n and RTDETR, enabling deployment across diverse AI devices with varying computational capabilities.</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

CAC-DETR: bridging heavy and light for real-time small object detection on UAVs

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zhixing Zhao,
  • Yudi Zhao,
  • Zhensong Li,
  • Zhihai Zhuo,
  • Rui Yin,
  • Kai Zhao

摘要

Real-time small object detection in unmanned aerial vehicle imagery serves critical applications but faces significant deployment challenges. The primary constraint lies in achieving optimal trade-offs between detection accuracy and computational efficiency on resource-limited edge-AI devices. To address small object detection accuracy while maintaining computational efficiency, we propose Convolutional Additive Context DEtection TRansformer (CAC-DETR), a novel detection framework based on Real-Time DEtection TRansformer (RT-DETR). Our framework introduces the CAC block, which comprises three core components: C2f, Convolutional Additive Token Mixer (CATM), and Convolutional Gated Linear Unit (CGLU). The diverse kernel sizes employed in CATM and CGLU provide flexible receptive field adjustment, while the pure convolutional architecture ensures optimal computational efficiency for edge deployment. Our proposed CAC-DETR achieves substantial performance improvements while reducing computational overhead compared to the benchmark RT-DETR. Experimental results on VisDrone demonstrate computational reductions of 30.3% in parameter complexity and 20.4% in GFLOPs, accompanied by significant accuracy improvements of 3.6% \(\text {mAP}_{0.5}\) mAP 0.5 and 4.0% \(\text {mAP}_{0.5:0.95}\) mAP 0.5 : 0.95 . Consistent performance gains are also observed on HIT-UAV and DIOR datasets. Furthermore, we employ Localization Knowledge Distillation to enhance the detection accuracy of lightweight models, such as YOLOv8-n and RTDETR, enabling deployment across diverse AI devices with varying computational capabilities.