Scholarlink

<p>With the wide application of Unmanned Aerial Vehicles (UAVs), the issue of missed and false detection of small objects in aerial photography needs to be solved. Most current works on small object detection either utilize more complex backbone networks or incorporate multiple manually designed modules within the feature fusion network. These approaches lead to an increase in the computational cost of the model while accuracy improvement is limited. To address the aforementioned challenges, this paper proposes FES-DETR, an efficient DEtection TRansformer (DETR) framework tailored for UAV imagery. Firstly, a multi-scale feature fusion method with frequency-domain enhancement across the spatial domain is proposed to enhance feature extraction by combining high-frequency feature information of imagery. Then, a lightweight low-loss downsampling module is proposed to avoid loss of feature information during image compression. Moreover, a small object detector head for high-resolution feature maps is added to improve the detection performance of small-scale features. Finally, an Inner-WIoU loss function is used to enhance convergence speed and accuracy. Experimental results indicate that FES-DETR achieves mAP@0.5 of 54.7% and mAP@0.5:0.95 of 34.5% on VisDrone, an 8.1% and 6.1% improvement over the baseline model, respectively.</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

FES-DETR: efficient UAV object detector with frequency-domain enhancement across spatial-domain

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Rong Qian,
  • Congzhe You

摘要

With the wide application of Unmanned Aerial Vehicles (UAVs), the issue of missed and false detection of small objects in aerial photography needs to be solved. Most current works on small object detection either utilize more complex backbone networks or incorporate multiple manually designed modules within the feature fusion network. These approaches lead to an increase in the computational cost of the model while accuracy improvement is limited. To address the aforementioned challenges, this paper proposes FES-DETR, an efficient DEtection TRansformer (DETR) framework tailored for UAV imagery. Firstly, a multi-scale feature fusion method with frequency-domain enhancement across the spatial domain is proposed to enhance feature extraction by combining high-frequency feature information of imagery. Then, a lightweight low-loss downsampling module is proposed to avoid loss of feature information during image compression. Moreover, a small object detector head for high-resolution feature maps is added to improve the detection performance of small-scale features. Finally, an Inner-WIoU loss function is used to enhance convergence speed and accuracy. Experimental results indicate that FES-DETR achieves mAP@0.5 of 54.7% and mAP@0.5:0.95 of 34.5% on VisDrone, an 8.1% and 6.1% improvement over the baseline model, respectively.