Scholarlink

<p>On resource-constrained platforms, small object detection in Unmanned Aerial Vehicle (UAV) images faces significant challenges due to the requirements of high-resolution input and real-time processing. We propose a real-time UAV aerial images small object detection framework, RT-UAV-SOD, which is based on the Transformer architecture and aims to improve the detection performance of small objects in UAV aerial images. In order to improve the detection performance, the model introduces two key innovations. First, we embed the Cascade Group Attention mechanism (CGA) into the inverted residual structure to construct the Cascade Group Attention-inverted Residual Mobile Block (CGA-iRMB) and form the backbone network. The optimized residual design strengthens the feature expression ability of the backbone network to enhance multi-scale feature extraction and to reduce computational complexity. Then, the cross-stage fusion module improves the multi-scale feature fusion process to achieve a balance between detection accuracy and inference speed. These improvements not only ensure the model’s compatibility with edge devices but also enhance the robust processing capability for aerial images. Experiments on the VisDrone2019-DET dataset demonstrate that our model increases precision by 3.3%, improves mAP50 by 4.5%, and enhances mAP50:95 by 2.5%. Experiments on the DOTA dataset demonstrate that our model increases precision by 4.3%, improves mAP50 by 1.1%, and enhances mAP50:95 by 2.6%. These results confirm that RT-UAV-SOD delivers an efficient solution for real-time object detection in UAV applications.</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

A real-time and efficient detector for small object in UAV aerial images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月10日

查看全文
  • Li Tan,
  • Chen Zhang,
  • Hua Bai,
  • Zikang Liu,
  • Yibo Li

摘要

On resource-constrained platforms, small object detection in Unmanned Aerial Vehicle (UAV) images faces significant challenges due to the requirements of high-resolution input and real-time processing. We propose a real-time UAV aerial images small object detection framework, RT-UAV-SOD, which is based on the Transformer architecture and aims to improve the detection performance of small objects in UAV aerial images. In order to improve the detection performance, the model introduces two key innovations. First, we embed the Cascade Group Attention mechanism (CGA) into the inverted residual structure to construct the Cascade Group Attention-inverted Residual Mobile Block (CGA-iRMB) and form the backbone network. The optimized residual design strengthens the feature expression ability of the backbone network to enhance multi-scale feature extraction and to reduce computational complexity. Then, the cross-stage fusion module improves the multi-scale feature fusion process to achieve a balance between detection accuracy and inference speed. These improvements not only ensure the model’s compatibility with edge devices but also enhance the robust processing capability for aerial images. Experiments on the VisDrone2019-DET dataset demonstrate that our model increases precision by 3.3%, improves mAP50 by 4.5%, and enhances mAP50:95 by 2.5%. Experiments on the DOTA dataset demonstrate that our model increases precision by 4.3%, improves mAP50 by 1.1%, and enhances mAP50:95 by 2.6%. These results confirm that RT-UAV-SOD delivers an efficient solution for real-time object detection in UAV applications.