Scholarlink

<p>Drone object detection is widely applied in various scenarios. However, drone aerial imagery suffers from complex backgrounds and small objects, leading to low recognition accuracy of detection models, and practical deployment requires strict real-time inference to ensure that detection results can be applied instantly in dynamic environments. To address these challenges, this paper proposes Drone DETR, a real-time end-to-end detection model for drone small object imagery based on the RT-DETR framework. Drone DETR incorporates a lightweight EfficientNet backbone enhanced with the designed spatial–channel–pixel attention module to achieve a better balance between accuracy and efficiency. In addition, a channel adaptive feature fusion module is developed to improve multi-scale feature integration by computing adaptive importance weights of feature elements. Furthermore, a small object feature pyramid network is constructed based on the proposed multi-scale feature fusion module, effectively compensating for the loss of small object details during downsampling. Moreover, a novel NWD-SIoU loss&#xa0;function is introduced to enhance bounding-box regression and improve the model’s performance in small object detection. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method improves mAP50 by 4.7% and mAP50:95 by 3.6% compared with the baseline. The model also achieves performance gains on the HIT-UAV dataset, confirming its effectiveness in addressing both accuracy and strict real-time requirements of drone small object detection.</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

Drone DETR: enhancing real-time detection transformer for drone small object imagery

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Weixi Wang,
  • Ye Liu,
  • Zhouzhou Cui,
  • Liang Shan,
  • Bo-Chao Zheng,
  • Zhong Wei,
  • Xiaoqing Zhang,
  • Satoshi Yamane

摘要

Drone object detection is widely applied in various scenarios. However, drone aerial imagery suffers from complex backgrounds and small objects, leading to low recognition accuracy of detection models, and practical deployment requires strict real-time inference to ensure that detection results can be applied instantly in dynamic environments. To address these challenges, this paper proposes Drone DETR, a real-time end-to-end detection model for drone small object imagery based on the RT-DETR framework. Drone DETR incorporates a lightweight EfficientNet backbone enhanced with the designed spatial–channel–pixel attention module to achieve a better balance between accuracy and efficiency. In addition, a channel adaptive feature fusion module is developed to improve multi-scale feature integration by computing adaptive importance weights of feature elements. Furthermore, a small object feature pyramid network is constructed based on the proposed multi-scale feature fusion module, effectively compensating for the loss of small object details during downsampling. Moreover, a novel NWD-SIoU loss function is introduced to enhance bounding-box regression and improve the model’s performance in small object detection. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method improves mAP50 by 4.7% and mAP50:95 by 3.6% compared with the baseline. The model also achieves performance gains on the HIT-UAV dataset, confirming its effectiveness in addressing both accuracy and strict real-time requirements of drone small object detection.