Scholarlink

<p>Unmanned aerial vehicle (UAV) imagery plays a vital role across numerous applications. However, small-object detection performance often remains unsatisfactory due to low spatial resolution, large-scale variations, and complex backgrounds. To address the above issues, this paper proposes FT-YOLO, an algorithm designed for small-object detection in UAV imagery. First, we construct a feature extraction unit named C3K2-PF, which integrates the poly kernel inception block with the feature enhancement module to&#xa0;enhance fine-grained and directional features, improving the ability to represent small objects. Second, we incorporate the attention scale fusion strategy into the neck, enabling adaptive fusion of multi-level features. In addition, we designed a task-synchronized dynamic head, which employs a task decomposition mechanism to separate and optimize classification and regression tasks synchronously. This approach avoids conflicts and enhances detection accuracy. Finally, the Wise-IoU loss function is adopted to optimize bounding box regression and accelerate model convergence. Experimental results show that FT-YOLO achieves a mean average precision (mAP) of 37.8% and a real-time detection speed of 97.1 frames per second on the VisDrone2019 dataset. Compared to the baseline model YOLOv11, it improves mAP50 and mAP50–95 by 5.9% and 3.6%, respectively. Cross-dataset evaluation further demonstrates the model’s robustness and its ability to generalize to diverse scenarios.</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

FT-YOLO: a UAV-oriented small-object detection algorithm based on feature enhancement and task synchronization

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Bing Jin,
  • Zhimin Lu,
  • Lesheng He,
  • Pengfei Yu

摘要

Unmanned aerial vehicle (UAV) imagery plays a vital role across numerous applications. However, small-object detection performance often remains unsatisfactory due to low spatial resolution, large-scale variations, and complex backgrounds. To address the above issues, this paper proposes FT-YOLO, an algorithm designed for small-object detection in UAV imagery. First, we construct a feature extraction unit named C3K2-PF, which integrates the poly kernel inception block with the feature enhancement module to enhance fine-grained and directional features, improving the ability to represent small objects. Second, we incorporate the attention scale fusion strategy into the neck, enabling adaptive fusion of multi-level features. In addition, we designed a task-synchronized dynamic head, which employs a task decomposition mechanism to separate and optimize classification and regression tasks synchronously. This approach avoids conflicts and enhances detection accuracy. Finally, the Wise-IoU loss function is adopted to optimize bounding box regression and accelerate model convergence. Experimental results show that FT-YOLO achieves a mean average precision (mAP) of 37.8% and a real-time detection speed of 97.1 frames per second on the VisDrone2019 dataset. Compared to the baseline model YOLOv11, it improves mAP50 and mAP50–95 by 5.9% and 3.6%, respectively. Cross-dataset evaluation further demonstrates the model’s robustness and its ability to generalize to diverse scenarios.