Scholarlink

<p>Unmanned aerial vehicle (UAV) plays an important role in various areas. To improve small object detection accuracy in UAV images, a small object detection model based on YOLOv7 has been proposed. Firstly, a multi-scale feature fusion structure for extracting small object features was configured. Secondly, the dynamic head with scale attention was introduced to focus on shallow feature map residing small objects. Finally, the auxiliary anchor label allocation strategy was proposed to provide positive labels for small objects via expanding the Ground Truth (GT) and prior anchor areas. The experiments were carried out on the VisDrone2019 dataset and UAVDT dataset to verify that the proposed algorithm can achieve higher detection accuracy in UAV image detection. The experimental results show that the proposed algorithm has achieved that mAP0.5 upgraded by 6%, mAP0.5:0.95 upgraded by 6.5% on VisDrone2019 and mAP0.5 upgraded by 1.2%, mAP0.5:0.95 upgraded by 4.5% on UAVDT dataset.</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

Aams-yolo: a small object detection method for UAV capture scenes based on YOLOv7

  • 期刊论文

  • 发表日期：

    2025年4月28日

查看全文
  • Yue Liu,
  • Jiale Zhang,
  • Shimiao Liu,
  • Linjun Xu,
  • Yafei Wang

摘要

Unmanned aerial vehicle (UAV) plays an important role in various areas. To improve small object detection accuracy in UAV images, a small object detection model based on YOLOv7 has been proposed. Firstly, a multi-scale feature fusion structure for extracting small object features was configured. Secondly, the dynamic head with scale attention was introduced to focus on shallow feature map residing small objects. Finally, the auxiliary anchor label allocation strategy was proposed to provide positive labels for small objects via expanding the Ground Truth (GT) and prior anchor areas. The experiments were carried out on the VisDrone2019 dataset and UAVDT dataset to verify that the proposed algorithm can achieve higher detection accuracy in UAV image detection. The experimental results show that the proposed algorithm has achieved that mAP0.5 upgraded by 6%, mAP0.5:0.95 upgraded by 6.5% on VisDrone2019 and mAP0.5 upgraded by 1.2%, mAP0.5:0.95 upgraded by 4.5% on UAVDT dataset.