Scholarlink

<p>With the rapid advancement of unmanned aerial vehicle (UAV) technology, small object detection in aerial images has become an important research direction in computer vision. However, aerial scenes present significant challenges, including large-scale variations, complex backgrounds, and dense target distributions. These factors make it difficult for existing detection approaches to achieve optimal balance between accuracy and computational costs. To address these challenges, this paper proposes HyperEdge-DETR, an improved Real-Time Detection Transformer-based algorithm for UAV small object detection. First, we designed a Multi-scale Adaptive Enhancement Network (MAENet) backbone based on Cross Stage Partial (CSP) modules with edge enhancement mechanisms. Second, a Graph-based Feature Integration Network (GFIN) is introduced that establishes high-order associations among cross-scale features through hypergraph computation. Finally, a Progressive Attention Integration (PAI) hierarchical attention fusion module is presented to enhance detection head capabilities. Experimental results on the VisDrone2019 dataset show that HyperEdge-DETR, with only 19.8M parameters and 64.9G FLOPs, achieves 22.5% in AP and 13.5% in APs, which are 2.5% and 2.2% higher than RT-DETR, respectively. More experimental results and analysis on additional datasets indicate that the proposed model provides an effective solution for small target detection in drone 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

HyperEdge-DETR: hypergraph-enhanced edge-aware detection transformer for small object detection in UAV

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Cheng Shen,
  • Shusen Sun,
  • Haixiang Zhang,
  • Mingfeng Jiang

摘要

With the rapid advancement of unmanned aerial vehicle (UAV) technology, small object detection in aerial images has become an important research direction in computer vision. However, aerial scenes present significant challenges, including large-scale variations, complex backgrounds, and dense target distributions. These factors make it difficult for existing detection approaches to achieve optimal balance between accuracy and computational costs. To address these challenges, this paper proposes HyperEdge-DETR, an improved Real-Time Detection Transformer-based algorithm for UAV small object detection. First, we designed a Multi-scale Adaptive Enhancement Network (MAENet) backbone based on Cross Stage Partial (CSP) modules with edge enhancement mechanisms. Second, a Graph-based Feature Integration Network (GFIN) is introduced that establishes high-order associations among cross-scale features through hypergraph computation. Finally, a Progressive Attention Integration (PAI) hierarchical attention fusion module is presented to enhance detection head capabilities. Experimental results on the VisDrone2019 dataset show that HyperEdge-DETR, with only 19.8M parameters and 64.9G FLOPs, achieves 22.5% in AP and 13.5% in APs, which are 2.5% and 2.2% higher than RT-DETR, respectively. More experimental results and analysis on additional datasets indicate that the proposed model provides an effective solution for small target detection in drone applications.