Scholarlink

<p>To address the challenges of target detection in drone images, such as complex backgrounds, severe occlusion, dense small targets, and varying lighting conditions, we propose an end-to-end small target multi-scale feature extraction and fusion detection network, STFF-RTDETR. To more accurately extract target features from complex backgrounds, we first introduce a reparameterized dilation module (DR-Block). By reparameterizing convolutions, we expand the receptive field and improve local feature extraction, making the model more sensitive to multi-scale targets and enhancing its attention to dense targets in complex backgrounds, thereby capturing more details. Next, we combine the HiLo attention mechanism with the intra-scale feature interaction module to form the HOAIFI feature interaction module, which is integrated into the hybrid encoder to enhance the model’s attention to high-frequency object information. Finally, we employ a cross-scale feature fusion pyramid network (P2-CCPF), where a P2 layer is added during the cross-scale feature fusion process to obtain rich small target information. SPDConv is used to preserve small target details while reducing computational complexity, achieving efficient cross-scale feature fusion. Experimental results show that compared to the RT-DETR baseline model, STFF-RTDETR improves the mAP0.5 score by 3.5% and 2.1% on the VisDrone and HIT-UAV datasets, 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

Stff-rtdetr: a small object detection algorithm based on drone aerial photography

  • 期刊论文

  • 发表日期：

    2025年5月30日

查看全文
  • Xiaoxi Teng,
  • Wendong Zhang,
  • Tao Liu,
  • Jingdong Yang,
  • Mengnan Ma

摘要

To address the challenges of target detection in drone images, such as complex backgrounds, severe occlusion, dense small targets, and varying lighting conditions, we propose an end-to-end small target multi-scale feature extraction and fusion detection network, STFF-RTDETR. To more accurately extract target features from complex backgrounds, we first introduce a reparameterized dilation module (DR-Block). By reparameterizing convolutions, we expand the receptive field and improve local feature extraction, making the model more sensitive to multi-scale targets and enhancing its attention to dense targets in complex backgrounds, thereby capturing more details. Next, we combine the HiLo attention mechanism with the intra-scale feature interaction module to form the HOAIFI feature interaction module, which is integrated into the hybrid encoder to enhance the model’s attention to high-frequency object information. Finally, we employ a cross-scale feature fusion pyramid network (P2-CCPF), where a P2 layer is added during the cross-scale feature fusion process to obtain rich small target information. SPDConv is used to preserve small target details while reducing computational complexity, achieving efficient cross-scale feature fusion. Experimental results show that compared to the RT-DETR baseline model, STFF-RTDETR improves the mAP0.5 score by 3.5% and 2.1% on the VisDrone and HIT-UAV datasets, respectively.