Scholarlink

To address the limitations in small object detection, complex scene understanding, and multi-scale object detection for UAVs, we propose a Dynamic Full-domain Feature Extraction Network (DF-DETR), an enhanced version optimized based on RT-DETR. Firstly, a Convolutional Dynamic Integration Module based on CSPNet (CDIMB) is introduced into the backbone. This module facilitates adaptive deep convolution and efficient feature fusion. Secondly, we introduce and design a Full-domain Transformer to enhance the AIFI module. Leveraging the Transformer architecture’s heightened sensitivity to low-frequency information, this modification significantly improves feature extraction capabilities. Furthermore, a Global-Local Spatial Attention Mechanism (GLSA) is embedded into the P3, P4, and P5 layers to strengthen attention to critical regions and suppress irrelevant background interference. On the VisDrone2019 dataset, the enhanced model achieves a 1.73% improvement in mAP50, while reducing the number of parameters by 22% and computational cost by 6.2%. Compared with other models of similar complexity, the proposed method demonstrates superior performance, validating the effectiveness of these enhancements.

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

DF-DETR: An Algorithm for Small Object Detection in UAV Aerial Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Liangyao Bu,
  • Hui Yang,
  • Bo Yang

摘要

To address the limitations in small object detection, complex scene understanding, and multi-scale object detection for UAVs, we propose a Dynamic Full-domain Feature Extraction Network (DF-DETR), an enhanced version optimized based on RT-DETR. Firstly, a Convolutional Dynamic Integration Module based on CSPNet (CDIMB) is introduced into the backbone. This module facilitates adaptive deep convolution and efficient feature fusion. Secondly, we introduce and design a Full-domain Transformer to enhance the AIFI module. Leveraging the Transformer architecture’s heightened sensitivity to low-frequency information, this modification significantly improves feature extraction capabilities. Furthermore, a Global-Local Spatial Attention Mechanism (GLSA) is embedded into the P3, P4, and P5 layers to strengthen attention to critical regions and suppress irrelevant background interference. On the VisDrone2019 dataset, the enhanced model achieves a 1.73% improvement in mAP50, while reducing the number of parameters by 22% and computational cost by 6.2%. Compared with other models of similar complexity, the proposed method demonstrates superior performance, validating the effectiveness of these enhancements.