Scholarlink

<p>Underwater object detection is severely hindered by image degradation from light absorption and scattering, challenging real-time applications. This paper introduces SCS-Mamba, a lightweight and efficient detection framework designed for resource-constrained underwater platforms. Our model integrates a Sobel-enhanced stem for domain-specific edge recovery, content-aware CARAFE upsampling for detail preservation, and a parameter-efficient Shared Deformable Adaptive Head into a Mamba-based backbone. The proposed architecture has only 1.3M parameters and 6.2 GFLOPs, enabling real-time performance. Extensive experiments on six public datasets show that SCS-Mamba achieves a new state-of-the-art for efficient underwater detection, with an average of 80.6% mAP@0.5. It consistently surpasses not only lightweight YOLO variants but also heavyweight general-purpose detectors and specialized underwater models. Its domain-specific design ensures robust performance, making it highly suitable for applications on autonomous underwater vehicles and marine monitoring systems.</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

SCS-Mamba: a lightweight Sobel-enhanced content-aware state space model with shared deformable adaptive head for efficient underwater object detection

  • 期刊论文

  • 发表日期：

    2025年11月9日

查看全文
  • Yili Xu,
  • Xueqi Zhao,
  • Xuanxuan Xiao

摘要

Underwater object detection is severely hindered by image degradation from light absorption and scattering, challenging real-time applications. This paper introduces SCS-Mamba, a lightweight and efficient detection framework designed for resource-constrained underwater platforms. Our model integrates a Sobel-enhanced stem for domain-specific edge recovery, content-aware CARAFE upsampling for detail preservation, and a parameter-efficient Shared Deformable Adaptive Head into a Mamba-based backbone. The proposed architecture has only 1.3M parameters and 6.2 GFLOPs, enabling real-time performance. Extensive experiments on six public datasets show that SCS-Mamba achieves a new state-of-the-art for efficient underwater detection, with an average of 80.6% mAP@0.5. It consistently surpasses not only lightweight YOLO variants but also heavyweight general-purpose detectors and specialized underwater models. Its domain-specific design ensures robust performance, making it highly suitable for applications on autonomous underwater vehicles and marine monitoring systems.