Scholarlink

<p>Uncrewed aerial vehicle (UAV) object detection remains challenging due to extremely small targets, large scale variations, and dense object distributions. To address the limitations of existing detectors in multi-scale representation and global context modeling, we propose the Dual Pyramid Attention Network (DPA-Net), which achieves real-time detection through coordinated enhancement of feature extraction and feature fusion. DPA-Net integrates three lightweight modules: (1) the Dual-Scale Patch Attention Module (DPAM) for fine-grained small object perception, (2) the Large-Kernel Spatial Pyramid Pooling (LKSPP) module for efficient long-range spatial context, and (3) the Channel-Spatial One-Shot Aggregation (CSOSA) module for cross-scale feature aggregation. Experiments on the VisDrone-DET dataset show that DPA-Net reaches 14.53% mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50:95}\)</EquationSource> </InlineEquation>, surpassing YOLOv8n by 1.19 points while maintaining real-time performance at 227.9 FPS, confirming its effectiveness for UAV scenarios.</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

DPA-Net: a real-time dual-pyramid attention network for UAV object detection

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Yang Cui,
  • Jianxun Shi,
  • Dong Guo

摘要

Uncrewed aerial vehicle (UAV) object detection remains challenging due to extremely small targets, large scale variations, and dense object distributions. To address the limitations of existing detectors in multi-scale representation and global context modeling, we propose the Dual Pyramid Attention Network (DPA-Net), which achieves real-time detection through coordinated enhancement of feature extraction and feature fusion. DPA-Net integrates three lightweight modules: (1) the Dual-Scale Patch Attention Module (DPAM) for fine-grained small object perception, (2) the Large-Kernel Spatial Pyramid Pooling (LKSPP) module for efficient long-range spatial context, and (3) the Channel-Spatial One-Shot Aggregation (CSOSA) module for cross-scale feature aggregation. Experiments on the VisDrone-DET dataset show that DPA-Net reaches 14.53% mAP \(_{50:95}\) , surpassing YOLOv8n by 1.19 points while maintaining real-time performance at 227.9 FPS, confirming its effectiveness for UAV scenarios.