Scholarlink

<p>Ship detection in complex maritime environments is primarily hindered by insufficient feature discrimination under dynamic background noise (e.g., sea glint and wave clutter) and high computational redundancy on edge hardware. This paper proposes MBF-DETR (multi-adaptive backbone feature detection Transformer), a lightweight real-time detector derived from RT-DETR. Three key enhancements are integrated: (1) MBB-HAF, an asymmetric dual-branch backbone that decouples semantic learning and structural extraction to improve object-to-background contrast; (2) MFM, which employs an adaptive weight learning mechanism to calibrate scale-wise importance, thereby mitigating the loss of small-vessel features in deep layers; and (3) MSAIFI, which utilizes parallel dilated convolutions to filter maritime clutter via context-aware self-attention. On the ZS-Ship dataset, MBF-DETR achieves 93.6% mAP@50 and 81.2% mAP@50:95 with only 11.8M parameters, outperforming the RT-DETR-r18 baseline by 4.4 and 5.5 percentage points, respectively. When deployed on an NVIDIA Jetson Xavier NX, the model maintains a stable inference rate of 33.7 FPS, proving its efficiency for intelligent maritime surveillance without compromising accuracy.</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

MBF-DETR: a lightweight real-time detector with multi-backbone hybrid attention for ship detection

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Xiaobing Fan,
  • Bowen Xing,
  • Xingchen Wang,
  • Hongdan Liu,
  • Chuanxu Yan,
  • Pengfei Zhi

摘要

Ship detection in complex maritime environments is primarily hindered by insufficient feature discrimination under dynamic background noise (e.g., sea glint and wave clutter) and high computational redundancy on edge hardware. This paper proposes MBF-DETR (multi-adaptive backbone feature detection Transformer), a lightweight real-time detector derived from RT-DETR. Three key enhancements are integrated: (1) MBB-HAF, an asymmetric dual-branch backbone that decouples semantic learning and structural extraction to improve object-to-background contrast; (2) MFM, which employs an adaptive weight learning mechanism to calibrate scale-wise importance, thereby mitigating the loss of small-vessel features in deep layers; and (3) MSAIFI, which utilizes parallel dilated convolutions to filter maritime clutter via context-aware self-attention. On the ZS-Ship dataset, MBF-DETR achieves 93.6% mAP@50 and 81.2% mAP@50:95 with only 11.8M parameters, outperforming the RT-DETR-r18 baseline by 4.4 and 5.5 percentage points, respectively. When deployed on an NVIDIA Jetson Xavier NX, the model maintains a stable inference rate of 33.7 FPS, proving its efficiency for intelligent maritime surveillance without compromising accuracy.