Scholarlink

<p>Unmanned aerial vehicle (UAV) imagery presents challenges including small object size, significant geometric deformations, and background clutter due to unique perspectives and complex backgrounds. To address these challenges while adapting to resource-constrained UAV platforms, this paper proposes MSDC-DETR, a real-time object detection model. The multi-scale deformable channel convolution (MSDC) module decomposes multi-scale features and dynamically adjusts channel weights to generate precise sampling offsets, enabling convolutional kernels to adaptively focus on object regions for enhanced spatial flexibility and mitigated geometric deformations. The adaptive fusion attention module bridges cross-scale semantic gaps through interactive alignment, effectively suppressing background noise and boosting small object detection while reducing computational overhead and memory consumption. Extensive experiments on VisDrone and UAVDT datasets demonstrate 3.8% mAP50 and 2.5% mAP50:95 improvements while sustaining 60.1 FPS real-time inference, mitigating accuracy-speed trade-offs to provide a perception solution with improved efficiency-accuracy balance for most UAV systems. The source code is available at: <a href="https://github.com/13849446299/MSDC-DETR.">https://github.com/13849446299/MSDC-DETR.</a></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 object detection through multi-scale deformable convolutions and adaptive fusion attention

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Xuebin Xu,
  • Ziyang Xing,
  • Meiling Sun,
  • Peiran Zhang,
  • Kuihe Yang

摘要

Unmanned aerial vehicle (UAV) imagery presents challenges including small object size, significant geometric deformations, and background clutter due to unique perspectives and complex backgrounds. To address these challenges while adapting to resource-constrained UAV platforms, this paper proposes MSDC-DETR, a real-time object detection model. The multi-scale deformable channel convolution (MSDC) module decomposes multi-scale features and dynamically adjusts channel weights to generate precise sampling offsets, enabling convolutional kernels to adaptively focus on object regions for enhanced spatial flexibility and mitigated geometric deformations. The adaptive fusion attention module bridges cross-scale semantic gaps through interactive alignment, effectively suppressing background noise and boosting small object detection while reducing computational overhead and memory consumption. Extensive experiments on VisDrone and UAVDT datasets demonstrate 3.8% mAP50 and 2.5% mAP50:95 improvements while sustaining 60.1 FPS real-time inference, mitigating accuracy-speed trade-offs to provide a perception solution with improved efficiency-accuracy balance for most UAV systems. The source code is available at: https://github.com/13849446299/MSDC-DETR.