Scholarlink

<p>Underwater object detection is vital for marine exploration and ecological monitoring. However, due to the low-quality original characteristics of underwater images, the diversity changes of underwater species, and the scarcity of samples, existing underwater object detection technologies struggle to address these challenges simultaneously. To this end, this paper proposes a class-incremental few-shot underwater object detection framework(CIFS-UD). First, we design an end-to-end multi-task architecture by integrating the existing underwater image enhancement module based on the deep inception and channel-wise attention (DICAM) and object detection, enhancing underwater visual quality and enabling more discriminative feature representations. Second, we develop a prototype-based multi-scale dual attention module (ProMSDA) that strengthens cross-scale feature correlations and key object regions to optimize class prototype representations. Finally, we propose a dynamic boundary-aware prototype collaboration optimization strategy (DB-PCO) that jointly constrains intra-class compactness and inter-class separability to alleviate catastrophic forgetting. Experiments on Brackish, Trashcan, and RUOD datasets verify that CIFS-UD can effectively adapt to novel classes and improve the class-incremental few-shot underwater object detection performance.The code and implementation are publicly available at <a href="https://github.com/snow102/Class-Incremental-Few-Shot-Underwater-Object-Detection-Framework">https://github.com/snow102/Class-Incremental-Few-Shot-Underwater-Object-Detection-Framework</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

Class-incremental few-shot underwater object detection framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月15日

查看全文
  • Xue Zhao,
  • Bin Zhou,
  • Yanjiang Wang

摘要

Underwater object detection is vital for marine exploration and ecological monitoring. However, due to the low-quality original characteristics of underwater images, the diversity changes of underwater species, and the scarcity of samples, existing underwater object detection technologies struggle to address these challenges simultaneously. To this end, this paper proposes a class-incremental few-shot underwater object detection framework(CIFS-UD). First, we design an end-to-end multi-task architecture by integrating the existing underwater image enhancement module based on the deep inception and channel-wise attention (DICAM) and object detection, enhancing underwater visual quality and enabling more discriminative feature representations. Second, we develop a prototype-based multi-scale dual attention module (ProMSDA) that strengthens cross-scale feature correlations and key object regions to optimize class prototype representations. Finally, we propose a dynamic boundary-aware prototype collaboration optimization strategy (DB-PCO) that jointly constrains intra-class compactness and inter-class separability to alleviate catastrophic forgetting. Experiments on Brackish, Trashcan, and RUOD datasets verify that CIFS-UD can effectively adapt to novel classes and improve the class-incremental few-shot underwater object detection performance.The code and implementation are publicly available at https://github.com/snow102/Class-Incremental-Few-Shot-Underwater-Object-Detection-Framework.