Scholarlink

<p>Detecting military vessels together with their onboard key components (radars and batteries) from visible-light imagery is challenging due to the extremely small size of components, strong scale imbalance between hulls and parts, and the stringent real-time constraints of guidance and edge deployment. Meanwhile, publicly available visible-light datasets for military vessels with fine-grained component annotations are scarce, which limits fair evaluation and practical development. To address these issues, we propose NF-DETR, a real-time end-to-end detector built upon RT-DETR, and construct a publicly available visible-light dataset, MShips-2025, containing destroyers, submarines, and aircraft carriers with annotations for both vessels and onboard components. NF-DETR introduces (1) LECoT, a local enhanced contextual Transformer block for stronger local feature extraction at lower cost, (2) MARSConv, a lightweight multi-branch reorganized convolution for detail-preserving downsampling and faster inference, and (3) a Focaler-ACDIoU loss that strengthens regression supervision for small targets. On MShips-2025, NF-DETR achieves 93.3% mAP@50 and 70.1% mAP@50:95, improving RT-DETR-L by +2.1% and +2.0%, respectively, with +1.1% recall. Meanwhile, it reduces parameters by 16.54% and GFLOPs by 8.9% and increases inference speed by 46.98% (43→63 FPS). Edge device experiments further confirm the practicality of NF-DETR for real-time applications.</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

NF-DETR: A real-time detection method for military vessels and onboard small targets

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Huan Liu,
  • Jinsheng Xu,
  • Zhenfei Lin,
  • Zongtao Guo,
  • Shuwen Wang

摘要

Detecting military vessels together with their onboard key components (radars and batteries) from visible-light imagery is challenging due to the extremely small size of components, strong scale imbalance between hulls and parts, and the stringent real-time constraints of guidance and edge deployment. Meanwhile, publicly available visible-light datasets for military vessels with fine-grained component annotations are scarce, which limits fair evaluation and practical development. To address these issues, we propose NF-DETR, a real-time end-to-end detector built upon RT-DETR, and construct a publicly available visible-light dataset, MShips-2025, containing destroyers, submarines, and aircraft carriers with annotations for both vessels and onboard components. NF-DETR introduces (1) LECoT, a local enhanced contextual Transformer block for stronger local feature extraction at lower cost, (2) MARSConv, a lightweight multi-branch reorganized convolution for detail-preserving downsampling and faster inference, and (3) a Focaler-ACDIoU loss that strengthens regression supervision for small targets. On MShips-2025, NF-DETR achieves 93.3% mAP@50 and 70.1% mAP@50:95, improving RT-DETR-L by +2.1% and +2.0%, respectively, with +1.1% recall. Meanwhile, it reduces parameters by 16.54% and GFLOPs by 8.9% and increases inference speed by 46.98% (43→63 FPS). Edge device experiments further confirm the practicality of NF-DETR for real-time applications.