Scholarlink

With the increasing use of UAVs, there has been growing interest in object detection within aerial imagery. However, challenges such as scale variation, the presence of densely packed small objects, and the limitations of existing algorithms in capturing fine-grained features persist. This paper proposes an enhanced YOLOv8-based method featuring a branch-weighted auxiliary fusion feature pyramid network and a specialized layer for detecting small objects. The proposed PKBFormer module, leveraging parallel depthwise convolutions, further improves multiscale contextual understanding. A balance factor and an exponential normalized Gaussian Wasserstein distance loss function are introduced to reduce positional sensitivity for small objects. Based on the VisDrone2019 dataset, the proposed PBE-YOLO achieves significant performance gains, increasing mAP50 from 38.5% to 48.3% and mAP50:95 from 22.9% to 29.3%. On the AI-TOD dataset, mAP50 increases by 7.3%, demonstrating the method’s effectiveness for small object detection.

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

PBE-YOLO: Small Object Detection Model of UAV Aerial Photography Based on Branch Weighted Fusion

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jia Wen,
  • Xinyi Wang,
  • Ting Zhang,
  • Jialin Li

摘要

With the increasing use of UAVs, there has been growing interest in object detection within aerial imagery. However, challenges such as scale variation, the presence of densely packed small objects, and the limitations of existing algorithms in capturing fine-grained features persist. This paper proposes an enhanced YOLOv8-based method featuring a branch-weighted auxiliary fusion feature pyramid network and a specialized layer for detecting small objects. The proposed PKBFormer module, leveraging parallel depthwise convolutions, further improves multiscale contextual understanding. A balance factor and an exponential normalized Gaussian Wasserstein distance loss function are introduced to reduce positional sensitivity for small objects. Based on the VisDrone2019 dataset, the proposed PBE-YOLO achieves significant performance gains, increasing mAP50 from 38.5% to 48.3% and mAP50:95 from 22.9% to 29.3%. On the AI-TOD dataset, mAP50 increases by 7.3%, demonstrating the method’s effectiveness for small object detection.