Scholarlink

<p>Integrating object detection with unmanned aerial vehicles (UAVs) has greatly improved the efficiency of real-time floating debris monitoring. However, factors such as complex water scenarios, dense small objects, and limited computational resources still cause major challenges. In response, this paper proposes FDU-YOLO, a model specifically designed for UAV surface floating debris detection. First, the Efficient Inverted Residual (EIR) is proposed to replace the bottleneck in C2f, enhancing feature extraction and addressing the problem of small object feature sparsity. Second, the Efficient Detect Head (EDHead) is proposed to alleviate feature loss caused by repeated channel compression operations. Additionally, the Dimension-Aware Focusing-Diffusion Network (DAFDN) is proposed to strengthen cross-scale feature fusion, thereby effectively reducing false detections caused by water surface reflections and waves. Experimental results demonstrate that compared with YOLOv8n, our model achieves a 11.7% improvement in mean average precision on the UAV Floating Debris dataset, along with a 22.2% reduction in giga floating-point operations and a 21.9% increase in small object recall rate. At the same time, our model achieves 32 frames per second (FPS) on the RK3588 embedded device, meeting the requirements for real-time 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

FDU-YOLO: efficient surface floating debris detection model for UAV perspectives detection

  • 期刊论文

  • 发表日期：

    2025年11月24日

查看全文
  • Yier Yan,
  • Zhibin Liang,
  • Jiayu Li,
  • Tao Zou,
  • Changhong Liu

摘要

Integrating object detection with unmanned aerial vehicles (UAVs) has greatly improved the efficiency of real-time floating debris monitoring. However, factors such as complex water scenarios, dense small objects, and limited computational resources still cause major challenges. In response, this paper proposes FDU-YOLO, a model specifically designed for UAV surface floating debris detection. First, the Efficient Inverted Residual (EIR) is proposed to replace the bottleneck in C2f, enhancing feature extraction and addressing the problem of small object feature sparsity. Second, the Efficient Detect Head (EDHead) is proposed to alleviate feature loss caused by repeated channel compression operations. Additionally, the Dimension-Aware Focusing-Diffusion Network (DAFDN) is proposed to strengthen cross-scale feature fusion, thereby effectively reducing false detections caused by water surface reflections and waves. Experimental results demonstrate that compared with YOLOv8n, our model achieves a 11.7% improvement in mean average precision on the UAV Floating Debris dataset, along with a 22.2% reduction in giga floating-point operations and a 21.9% increase in small object recall rate. At the same time, our model achieves 32 frames per second (FPS) on the RK3588 embedded device, meeting the requirements for real-time detection.