Scholarlink

<p>Object detection technology has been widely applied in the field of remote sensing. However, the small size, dense distribution, and complex background of objects in remote sensing imagery make it difficult for existing detectors to achieve a favorable trade-off between accuracy and computational efficiency, particularly under resource-constrained deployment scenarios such as unmanned aerial systems. To address this issue, we propose a lightweight detection framework named QS-YOLOv8. First, we propose a high-level feature channel compression strategy, which compresses redundant high-level channels and removes the P5 detection head, thereby reducing model complexity while preserving key semantic information. A visual serial feature fusion module is designed to enhance multi-scale contextual interaction, and a channel-decoupled bottleneck is introduced to strengthen feature representation capability under limited computational budgets. In addition, the P2 high-resolution feature is incorporated into the feature pyramid fusion process. Experimental results on three public remote sensing datasets, including VisDrone2019, AI-TOD, and DOTA, demonstrate that QS-YOLOv8 achieves consistent performance gains while maintaining a lightweight model scale. On VisDrone2019, the proposed method requires only 3.0&#xa0;M parameters and 22.2 GFLOPs, while achieving a clear improvement in detection accuracy.</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

QS-YOLOv8: a lightweight and efficient detector for small objects

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Shujing Xie,
  • Haikun Li,
  • Weihao Ye,
  • Jikun Chen,
  • Anqi Xiong,
  • Peiran Zhang

摘要

Object detection technology has been widely applied in the field of remote sensing. However, the small size, dense distribution, and complex background of objects in remote sensing imagery make it difficult for existing detectors to achieve a favorable trade-off between accuracy and computational efficiency, particularly under resource-constrained deployment scenarios such as unmanned aerial systems. To address this issue, we propose a lightweight detection framework named QS-YOLOv8. First, we propose a high-level feature channel compression strategy, which compresses redundant high-level channels and removes the P5 detection head, thereby reducing model complexity while preserving key semantic information. A visual serial feature fusion module is designed to enhance multi-scale contextual interaction, and a channel-decoupled bottleneck is introduced to strengthen feature representation capability under limited computational budgets. In addition, the P2 high-resolution feature is incorporated into the feature pyramid fusion process. Experimental results on three public remote sensing datasets, including VisDrone2019, AI-TOD, and DOTA, demonstrate that QS-YOLOv8 achieves consistent performance gains while maintaining a lightweight model scale. On VisDrone2019, the proposed method requires only 3.0 M parameters and 22.2 GFLOPs, while achieving a clear improvement in detection accuracy.