Scholarlink

<p>Aerial object detection is one of the fundamental tasks in applications of Unmanned aerial vehicle (UAV). However, aerial object has weak appearance due to long range and specific scene of photography. This paper proposes a multi-stage feature fusion network MFFNet based on Yolo framework for UAV Aerial object detection. Multi-stage feature fusion modules play the role of the neck in the whole network, consisting of an attention-based feature fusion module (A-FFM) for fusing cross-scale features, feature enhancement module (FEM) for fusing contextual feature each scale, and adaptive channel and spatial feature fusion module (ACSFF) for supporting each other. These fusing modules are helpful to aggregate features of candidate object regions. Meanwhile, we create the custom dataset AP-RB on water conservancy project. Experimental results on our custom dataset AP-RB show that the proposed MFFNet achieves 6.71%, 1.08%,0.42% mAP higher than baseline and SOTA models DDSCNet, RT-DETR, respectively. It verifies that the MFFNet is effective to handle objects hard to discriminate, and automatic monitoring and management are efficient by means of abnormal object detection in the practical scene of water conservancy project.</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

Multi-stage feature fusion network for UAV aerial object detection

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Xueqiang Zhao,
  • Xin Ma,
  • Wei Zhou

摘要

Aerial object detection is one of the fundamental tasks in applications of Unmanned aerial vehicle (UAV). However, aerial object has weak appearance due to long range and specific scene of photography. This paper proposes a multi-stage feature fusion network MFFNet based on Yolo framework for UAV Aerial object detection. Multi-stage feature fusion modules play the role of the neck in the whole network, consisting of an attention-based feature fusion module (A-FFM) for fusing cross-scale features, feature enhancement module (FEM) for fusing contextual feature each scale, and adaptive channel and spatial feature fusion module (ACSFF) for supporting each other. These fusing modules are helpful to aggregate features of candidate object regions. Meanwhile, we create the custom dataset AP-RB on water conservancy project. Experimental results on our custom dataset AP-RB show that the proposed MFFNet achieves 6.71%, 1.08%,0.42% mAP higher than baseline and SOTA models DDSCNet, RT-DETR, respectively. It verifies that the MFFNet is effective to handle objects hard to discriminate, and automatic monitoring and management are efficient by means of abnormal object detection in the practical scene of water conservancy project.