Scholarlink

<p>Underwater object detection is one of great significance in marine conservation, resource exploration, and scientific research. However, current underwater object detection faces two major challenges: first, the detection accuracy is not high due to complex backgrounds and low visibility; second, the model has a large number of parameters and severe memory consumption. Achieving a balance between high-precision detection and low parameter volume is an important task. This paper presents DMFI-YOLO, a real-time detection network based on YOLOv8, tailored for underwater scenes. DMFI-YOLO proposes a novel RFD-DarkNet backbone to enhance feature extraction capabilities, proposes a Gather-and-Distribute mechanism for multi-scale feature fusion, and proposes a dynamic task-aligned head (DT-Head) to improve localization and classification performance. Experiments on multiple datasets demonstrate that DMFI-YOLO achieves state-of-the-art results, improving mAP0.5:0.95 by 3.9% compared to YOLOv8-S on the DUO dataset, and showcases its effectiveness in detecting small and densely distributed marine organisms. The source code is publicly available at <a href="https://github.com/Lucky-954/DMFI-YOLO">https://github.com/Lucky-954/DMFI-YOLO</a>.</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

DMFI-YOLO: dynamic multi-scale feature interaction for enhanced underwater object detection based on YOLO

  • 期刊论文

  • 发表日期：

    2025年5月20日

查看全文
  • Xueyu Yu,
  • Yong Liu,
  • Hao Hu,
  • Xinzhi Li,
  • Mingdi Bo,
  • Dong Zhang,
  • Zijun Zhou

摘要

Underwater object detection is one of great significance in marine conservation, resource exploration, and scientific research. However, current underwater object detection faces two major challenges: first, the detection accuracy is not high due to complex backgrounds and low visibility; second, the model has a large number of parameters and severe memory consumption. Achieving a balance between high-precision detection and low parameter volume is an important task. This paper presents DMFI-YOLO, a real-time detection network based on YOLOv8, tailored for underwater scenes. DMFI-YOLO proposes a novel RFD-DarkNet backbone to enhance feature extraction capabilities, proposes a Gather-and-Distribute mechanism for multi-scale feature fusion, and proposes a dynamic task-aligned head (DT-Head) to improve localization and classification performance. Experiments on multiple datasets demonstrate that DMFI-YOLO achieves state-of-the-art results, improving mAP0.5:0.95 by 3.9% compared to YOLOv8-S on the DUO dataset, and showcases its effectiveness in detecting small and densely distributed marine organisms. The source code is publicly available at https://github.com/Lucky-954/DMFI-YOLO.