Scholarlink

<p>Small objects in aerial imagery often exhibit low resolution, weak signals, and limited semantic cues, while repeated downsampling in existing deep learning-based detectors further degrades critical spatial details. To balance detection accuracy and computational efficiency, we propose MRFNet, a <b>M</b>ulti-Scale <b>R</b>eparameterized <b>F</b>usion <b>Net</b>work built upon the YOLOv11 architecture. Specifically, we design a <b>T</b>riple-path <b>R</b>eparameterization <b>C</b>hannel (TRC) module that utilizes multi-scale receptive fields during training to enhance the semantic representation of small objects. To address detail loss in feature fusion, a <b>M</b>ulti-<b>S</b>cale <b>P</b>roduct <b>F</b>usion (MSPF) module is designed for the bottom-up path, while an <b>A</b>daptive <b>W</b>eighed <b>F</b>usion (AWF) module is incorporated into the top-down path to emphasize informative features. The detection head replaces the conventional low-resolution P5 layer with a high-resolution P2 layer, and the optimally configured windmill convolution is adopted to better capture directional structures of small objects. Experiments on the VisDrone2019, DIOR, and RSOD datasets achieve 47.7%, 79.0%, and 94.6% mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation>, surpassing YOLOv11n by 50.0%, 2.9%, and 0.3%, respectively. With fewer parameters and reduced computational cost, MRFNet achieves state-of-the-art performance and is successfully deployed on mobile edge devices, demonstrating its practical value for deployment on resource-constrained airborne platforms. The code is available at <a href="https://github.com/JSJ515-Group/MRFNet">https://github.com/JSJ515-Group/MRFNet</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

MRFNet: multi-scale reparameterized fusion network for efficient small object detection in aerial imagery

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Xingzhu Liang,
  • Jiale Xu,
  • Haifeng Xu,
  • Yu-e Lin,
  • Qicheng Hu,
  • Shanlin Shen

摘要

Small objects in aerial imagery often exhibit low resolution, weak signals, and limited semantic cues, while repeated downsampling in existing deep learning-based detectors further degrades critical spatial details. To balance detection accuracy and computational efficiency, we propose MRFNet, a Multi-Scale Reparameterized Fusion Network built upon the YOLOv11 architecture. Specifically, we design a Triple-path Reparameterization Channel (TRC) module that utilizes multi-scale receptive fields during training to enhance the semantic representation of small objects. To address detail loss in feature fusion, a Multi-Scale Product Fusion (MSPF) module is designed for the bottom-up path, while an Adaptive Weighed Fusion (AWF) module is incorporated into the top-down path to emphasize informative features. The detection head replaces the conventional low-resolution P5 layer with a high-resolution P2 layer, and the optimally configured windmill convolution is adopted to better capture directional structures of small objects. Experiments on the VisDrone2019, DIOR, and RSOD datasets achieve 47.7%, 79.0%, and 94.6% mAP \(_{50}\) 50 , surpassing YOLOv11n by 50.0%, 2.9%, and 0.3%, respectively. With fewer parameters and reduced computational cost, MRFNet achieves state-of-the-art performance and is successfully deployed on mobile edge devices, demonstrating its practical value for deployment on resource-constrained airborne platforms. The code is available at https://github.com/JSJ515-Group/MRFNet.