Scholarlink

<p>The rapid advancement and widespread adoption of unmanned aerial vehicles (UAVs) have spurred growing interest in aerial object detection. However, UAV-captured imagery presents significant challenges, including scale variations, cluttered backgrounds, object occlusions, and especially the presence of small objects. While transformer-based models like DETR show promise, they often fall short in extracting fine-grained features crucial for detecting small and densely packed objects. To overcome these challenges, we propose Polarized Attention and Multi-scale Fusion DEtection TRansformer (PAMF-DETR). PAMF-DETR features a lightweight backbone equipped with a Multi-Scale Feature Aggregation (MSFA) block to effectively extract fine-grained features of small objects. We propose a Scale-Compensated Cross-Channel Feature Fusion (SC-CCFF) structure that compensates for scale variations and enhances global context modeling by integrating dual-domain attention with large-kernel convolutions. Additionally, a polarity-aware attention mechanism is embedded into the hybrid encoder to improve spatial edge representation. For robust small-object localization, we adopt Inner-MPDIoU, which mitigates background interference during bounding box regression. Extensive experiments on the VisDrone dataset demonstrate that PAMF-DETR outperforms existing mainstream detectors in complex aerial scenarios, achieves a 1.9% improvement in <i>AP</i> and a 3.5% gain in <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(AP_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mn>50</mn> </msub> </mrow> </math></EquationSource> </InlineEquation>, while reducing the parameter count by 22.1% compared to the baseline. Similar improvements are observed on the DOTA dataset.</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

PAMF-DETR: A multi-scale-aware transformer for small object detection in UAV imagery

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Lingxuan Cao,
  • Jianming Wang,
  • Yukuan Sun,
  • Xiuyan Li

摘要

The rapid advancement and widespread adoption of unmanned aerial vehicles (UAVs) have spurred growing interest in aerial object detection. However, UAV-captured imagery presents significant challenges, including scale variations, cluttered backgrounds, object occlusions, and especially the presence of small objects. While transformer-based models like DETR show promise, they often fall short in extracting fine-grained features crucial for detecting small and densely packed objects. To overcome these challenges, we propose Polarized Attention and Multi-scale Fusion DEtection TRansformer (PAMF-DETR). PAMF-DETR features a lightweight backbone equipped with a Multi-Scale Feature Aggregation (MSFA) block to effectively extract fine-grained features of small objects. We propose a Scale-Compensated Cross-Channel Feature Fusion (SC-CCFF) structure that compensates for scale variations and enhances global context modeling by integrating dual-domain attention with large-kernel convolutions. Additionally, a polarity-aware attention mechanism is embedded into the hybrid encoder to improve spatial edge representation. For robust small-object localization, we adopt Inner-MPDIoU, which mitigates background interference during bounding box regression. Extensive experiments on the VisDrone dataset demonstrate that PAMF-DETR outperforms existing mainstream detectors in complex aerial scenarios, achieves a 1.9% improvement in AP and a 3.5% gain in \(AP_{50}\) A P 50 , while reducing the parameter count by 22.1% compared to the baseline. Similar improvements are observed on the DOTA dataset.