Scholarlink

<p>Complex underwater environments, characterized by blurred objects and strong background interference, present significant challenges for existing object detection algorithms. Accurate and efficient underwater object detection is crucial for sustainable marine resource development and conservation. This study presents the Dynamic Underwater Cognition Network (DUCNet), a lightweight yet high-accuracy model specifically designed for underwater small object detection in challenging conditions. To address underwater target blurring, we propose a Dynamic Channel-Aware Convolution (DCAConv) module that enhances feature extraction by integrating channel attention mechanisms with dynamic grouping strategies. We further develop a Dynamic Lightweight Feature Pyramid Network (DyLiteFPN) that employs dynamic kernel selection to improve multi-scale feature fusion efficiency, effectively tackling the scale variation challenges inherent in underwater imagery. Additionally, an Aligned Detection Head (AlignDet) is introduced, incorporating dynamic convolutional offsets and parameter-sharing techniques to resolve classification-localization misalignment issues in underwater object detection. Comprehensive experiments conducted on the DUO and RUOD datasets demonstrate that DUCNet achieves superior detection accuracy compared to state-of-the-art methods, with particularly notable improvements in small underwater object detection. Cross-dataset evaluations further validate the model’s excellent generalization capability and robustness. Extensive ablation studies confirm the effectiveness and necessity of each proposed module in enhancing overall model performance. The source code is available at <a href="https://github.com/lzbox/DUCNet">https://github.com/lzbox/DUCNet</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

Dynamic underwater cognition: aligned detection networks for enhanced underwater object recognition

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Dehua Zhang,
  • Zhen Li,
  • Changcheng Yu,
  • Ruixue Xia

摘要

Complex underwater environments, characterized by blurred objects and strong background interference, present significant challenges for existing object detection algorithms. Accurate and efficient underwater object detection is crucial for sustainable marine resource development and conservation. This study presents the Dynamic Underwater Cognition Network (DUCNet), a lightweight yet high-accuracy model specifically designed for underwater small object detection in challenging conditions. To address underwater target blurring, we propose a Dynamic Channel-Aware Convolution (DCAConv) module that enhances feature extraction by integrating channel attention mechanisms with dynamic grouping strategies. We further develop a Dynamic Lightweight Feature Pyramid Network (DyLiteFPN) that employs dynamic kernel selection to improve multi-scale feature fusion efficiency, effectively tackling the scale variation challenges inherent in underwater imagery. Additionally, an Aligned Detection Head (AlignDet) is introduced, incorporating dynamic convolutional offsets and parameter-sharing techniques to resolve classification-localization misalignment issues in underwater object detection. Comprehensive experiments conducted on the DUO and RUOD datasets demonstrate that DUCNet achieves superior detection accuracy compared to state-of-the-art methods, with particularly notable improvements in small underwater object detection. Cross-dataset evaluations further validate the model’s excellent generalization capability and robustness. Extensive ablation studies confirm the effectiveness and necessity of each proposed module in enhancing overall model performance. The source code is available at https://github.com/lzbox/DUCNet.