Scholarlink

<p>Accurate detection of small targets with extreme scale variations is crucial for unmanned aerial vehicle (UAV) imagery, yet it remains a significant challenge for traditional detectors. In this work, a novel and lightweight framework, LightS-DETR, is proposed. It integrates three innovative components: a selective context fusion network (SCFNet), a small-object-enhanced pyramid (SOEP), and an Inner-Shape IoU loss function. Together, these modules efficiently extract global context via bidirectional state-space modeling, prevent feature degradation during multiscale fusion, and dynamically adapt bounding box regression for irregular targets. The effectiveness of LightS-DETR was comprehensively evaluated on the RSOD and VisDrone2019 datasets. The proposed model achieves a highly competitive mean average precision of 98.36% on RSOD and 50.7% (mAP@0.5) on VisDrone2019. Maintaining a compact architecture of only 18.0 million parameters and 76.0 GFLOPs, LightS-DETR provides a real-time inference speed of 46.0 frames per second. Compared to existing state-of-the-art methods, LightS-DETR delivers superior detection accuracy and robust real-time performance while significantly reducing computational demands, offering an optimal balance for UAV deployment.</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

Improved detection accuracy and efficiency for small-object detection in drone imagery with lightweight LightS-DETR

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Xi Cai,
  • Shasha Zhao,
  • Changsuo Yu,
  • Dengying Zhang,
  • Xianwu Tang

摘要

Accurate detection of small targets with extreme scale variations is crucial for unmanned aerial vehicle (UAV) imagery, yet it remains a significant challenge for traditional detectors. In this work, a novel and lightweight framework, LightS-DETR, is proposed. It integrates three innovative components: a selective context fusion network (SCFNet), a small-object-enhanced pyramid (SOEP), and an Inner-Shape IoU loss function. Together, these modules efficiently extract global context via bidirectional state-space modeling, prevent feature degradation during multiscale fusion, and dynamically adapt bounding box regression for irregular targets. The effectiveness of LightS-DETR was comprehensively evaluated on the RSOD and VisDrone2019 datasets. The proposed model achieves a highly competitive mean average precision of 98.36% on RSOD and 50.7% (mAP@0.5) on VisDrone2019. Maintaining a compact architecture of only 18.0 million parameters and 76.0 GFLOPs, LightS-DETR provides a real-time inference speed of 46.0 frames per second. Compared to existing state-of-the-art methods, LightS-DETR delivers superior detection accuracy and robust real-time performance while significantly reducing computational demands, offering an optimal balance for UAV deployment.