Scholarlink

<p>To address critical challenges in UAV remote sensing small object detection–including weak feature representation, high background interference sensitivity, insufficient edge localization precision, and the accuracy–efficiency trade-off–this paper proposes YOLO11-MGNB, an enhanced algorithm based on the YOLO11 framework. The solution integrates four key innovations: (1) A multi-scale small object enhancement architecture is designed, which preserves fine-grained details of small targets through space-to-depth convolution transformation. Integrated with the adaptive multi-scale kernel fusion module, it achieves cross-domain feature interaction, enhancing micro-target representation while avoiding the high computational overhead of traditional P2 detection layers; (2) A global edge information propagator module enabling multi-level fusion of shallow edge features and deep semantics to enhance boundary sensitivity; (3) Normalized Wasserstein distance loss optimization to refine small target localization; (4) A bi-level routing attention mechanism constructing global–local collaborative screening to suppress complex background interference. Experimental results on the DIOR and VisDrone2019 datasets demonstrate that YOLO11-MGNB achieves highly competitive accuracy (87.7% and 32.0% mAP50, respectively) while maintaining ultra-lightweight characteristics of only 3.7 M parameters and 13.6 GFLOPs. More importantly, with an inference speed of 333.3 frames per second (FPS), it is considerably faster than other state-of-the-art models, e.g., 238.1 FPS higher than YOLO-MS and 272.4 FPS higher than SMA-YOLO, providing a superior solution for high-precision real-time detection in UAV remote sensing applications.</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

YOLO11-MGNB: lightweight real-time small object detection algorithm for UAV remote sensing images

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Weijie Ren,
  • Shuailong Zhang,
  • Zhike Zhou

摘要

To address critical challenges in UAV remote sensing small object detection–including weak feature representation, high background interference sensitivity, insufficient edge localization precision, and the accuracy–efficiency trade-off–this paper proposes YOLO11-MGNB, an enhanced algorithm based on the YOLO11 framework. The solution integrates four key innovations: (1) A multi-scale small object enhancement architecture is designed, which preserves fine-grained details of small targets through space-to-depth convolution transformation. Integrated with the adaptive multi-scale kernel fusion module, it achieves cross-domain feature interaction, enhancing micro-target representation while avoiding the high computational overhead of traditional P2 detection layers; (2) A global edge information propagator module enabling multi-level fusion of shallow edge features and deep semantics to enhance boundary sensitivity; (3) Normalized Wasserstein distance loss optimization to refine small target localization; (4) A bi-level routing attention mechanism constructing global–local collaborative screening to suppress complex background interference. Experimental results on the DIOR and VisDrone2019 datasets demonstrate that YOLO11-MGNB achieves highly competitive accuracy (87.7% and 32.0% mAP50, respectively) while maintaining ultra-lightweight characteristics of only 3.7 M parameters and 13.6 GFLOPs. More importantly, with an inference speed of 333.3 frames per second (FPS), it is considerably faster than other state-of-the-art models, e.g., 238.1 FPS higher than YOLO-MS and 272.4 FPS higher than SMA-YOLO, providing a superior solution for high-precision real-time detection in UAV remote sensing applications.