Scholarlink

This study addresses the challenges of aerial small-object detection in high-density urban environments, where heavy traffic and complex backgrounds often degrade detection accuracy, while practical applications also demand lightweight models. We propose PGS-YOLO, a precise and efficient detection framework. Specifically, a C3k2_PConv module is introduced into the backbone to selectively update key channels, reducing computation while preserving feature integrity. A Global-Local Spatial Attention Multiscale Feature Pyramid Network (GLMiFPN) is designed to enhance multi-scale feature fusion by jointly aggregating global context and refining local details. In addition, the SEAMHead detection head, equipped with channel and spatial attention, decouples classification and regression to improve information flow and accuracy without sacrificing efficiency. Furthermore, the Three-scale Feature Enhancement (TFE) module is refined to align and fuse P1–P3 features, enabling compact multi-scale prediction with a single head. Extensive experiments demonstrate that PGS-YOLO achieves mAP50 scores of 42.5% and 36.9% on the VisDrone-2019-DET and UAV-DT datasets, with improvements of 28% and 31.7% over the baseline, while adding only 0.4M parameters and 6.5 GFLOPs, confirming its effectiveness in balancing lightweight design and high accuracy under complex urban scenarios.

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

PGS-YOLO: A Lightweight and Accurate Framework for Aerial Small Object Detection in Urban Environments

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dongsheng Wang,
  • Yifei Wang,
  • Yuan Zhu

摘要

This study addresses the challenges of aerial small-object detection in high-density urban environments, where heavy traffic and complex backgrounds often degrade detection accuracy, while practical applications also demand lightweight models. We propose PGS-YOLO, a precise and efficient detection framework. Specifically, a C3k2_PConv module is introduced into the backbone to selectively update key channels, reducing computation while preserving feature integrity. A Global-Local Spatial Attention Multiscale Feature Pyramid Network (GLMiFPN) is designed to enhance multi-scale feature fusion by jointly aggregating global context and refining local details. In addition, the SEAMHead detection head, equipped with channel and spatial attention, decouples classification and regression to improve information flow and accuracy without sacrificing efficiency. Furthermore, the Three-scale Feature Enhancement (TFE) module is refined to align and fuse P1–P3 features, enabling compact multi-scale prediction with a single head. Extensive experiments demonstrate that PGS-YOLO achieves mAP50 scores of 42.5% and 36.9% on the VisDrone-2019-DET and UAV-DT datasets, with improvements of 28% and 31.7% over the baseline, while adding only 0.4M parameters and 6.5 GFLOPs, confirming its effectiveness in balancing lightweight design and high accuracy under complex urban scenarios.