Scholarlink

<p>To address the challenges of small object scales, dense object distributions, and complex background interference in unmanned aerial vehicle imagery, this paper proposes an improved small object detection method based on YOLOv11s. An enhanced multi-scale pooling fusion module is introduced to aggregate contextual information from different receptive fields and improve feature representation for small objects. An improved decoupled detection head is designed to separately optimize classification and localization, reducing task interference and enhancing localization accuracy for densely distributed targets. In addition, a same-scale spatial fusion module is proposed to strengthen spatial consistency among features at the same resolution, thereby improving small object recognition in complex scenes. Extensive experiments on the VisDrone2019 dataset demonstrate that the proposed method achieves an improvement of approximately 10 percentage points in mAP@50 over the baseline YOLOv11s, corresponding to a relative gain of about 27.5%, and exhibits improved robustness under challenging conditions such as motion blur and low illumination, highlighting its effectiveness for small object detection under a reasonable computational cost. The proposed method prioritizes detection accuracy with acceptable real-time performance.</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 YOLOv11s based on multi-scale pooling fusion for UAV imagery object detection

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Zihao Liu,
  • Yubin Fang,
  • Mengchu Tian,
  • Jie Chen,
  • Xuecheng Zhang

摘要

To address the challenges of small object scales, dense object distributions, and complex background interference in unmanned aerial vehicle imagery, this paper proposes an improved small object detection method based on YOLOv11s. An enhanced multi-scale pooling fusion module is introduced to aggregate contextual information from different receptive fields and improve feature representation for small objects. An improved decoupled detection head is designed to separately optimize classification and localization, reducing task interference and enhancing localization accuracy for densely distributed targets. In addition, a same-scale spatial fusion module is proposed to strengthen spatial consistency among features at the same resolution, thereby improving small object recognition in complex scenes. Extensive experiments on the VisDrone2019 dataset demonstrate that the proposed method achieves an improvement of approximately 10 percentage points in mAP@50 over the baseline YOLOv11s, corresponding to a relative gain of about 27.5%, and exhibits improved robustness under challenging conditions such as motion blur and low illumination, highlighting its effectiveness for small object detection under a reasonable computational cost. The proposed method prioritizes detection accuracy with acceptable real-time performance.