Scholarlink

<p>Tiny object detection remains a challenging task in computer vision, with broad applications in remote sensing, intelligent transportation, and aerial surveillance. Although recent advancements have improved detection accuracy, DETR-based methods such as RT-DETR still struggle with tiny objects due to limited receptive fields, aliasing artifacts, and the loss of fine-grained details. To address these challenges, we propose WT-DETR, an enhanced version of RT-DETR that incorporates wavelet transform-based optimizations. WT-DETR introduces Wave Field Convolution (WFC) to expand the receptive field while capturing global context and structural features with minimal parameter overhead. Furthermore, Wavelet Anti-Aliasing Downsampling (WTD) replaces conventional downsampling to mitigate aliasing and retain fine details, while maintaining computational efficiency. By integrating these components, WT-DETR improves multi-scale feature representation without sacrificing speed. Extensive experiments on the VisDrone2019 and SIMD datasets demonstrate that WT-DETR achieves <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1761_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="52" /> </InlineMediaObject> <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> scores of 59.65% and 81.0%, respectively, while maintaining an inference speed of 90.3 FPS–striking an effective balance between accuracy and real-time performance, and delivering competitive results compared to state-of-the-art methods.</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

WT-DETR: Wavelet-enhanced DETR for robust tiny object detection via multi-scale feature optimization

  • 期刊论文

  • 发表日期：

    2025年9月12日

查看全文
  • Xiaoyan Shao,
  • Shiqin Diao,
  • Lingling Li,
  • Xuezhuan Zhao,
  • Yang Mei,
  • Zonghao Zhu

摘要

Tiny object detection remains a challenging task in computer vision, with broad applications in remote sensing, intelligent transportation, and aerial surveillance. Although recent advancements have improved detection accuracy, DETR-based methods such as RT-DETR still struggle with tiny objects due to limited receptive fields, aliasing artifacts, and the loss of fine-grained details. To address these challenges, we propose WT-DETR, an enhanced version of RT-DETR that incorporates wavelet transform-based optimizations. WT-DETR introduces Wave Field Convolution (WFC) to expand the receptive field while capturing global context and structural features with minimal parameter overhead. Furthermore, Wavelet Anti-Aliasing Downsampling (WTD) replaces conventional downsampling to mitigate aliasing and retain fine details, while maintaining computational efficiency. By integrating these components, WT-DETR improves multi-scale feature representation without sacrificing speed. Extensive experiments on the VisDrone2019 and SIMD datasets demonstrate that WT-DETR achieves \(mAP_{50}\) m A P 50 scores of 59.65% and 81.0%, respectively, while maintaining an inference speed of 90.3 FPS–striking an effective balance between accuracy and real-time performance, and delivering competitive results compared to state-of-the-art methods.