Scholarlink

<p>Marine biofouling on ship hulls reduces navigation efficiency and increases maintenance costs, making accurate semantic segmentation critical for intelligent ship maintenance. Underwater imaging suffers from low contrast, uneven illumination, and large fouling scale variations, which degrade standard segmentation models. This work presents an enhanced UNeXt architecture named FMA-UNeXt for underwater hull fouling segmentation. The model integrates frequency-adaptive dilated convolution to enhance high-frequency details and low-frequency structures, a multi-scale convolutional attention decoder to handle scale variations, and efficient multi-scale attention modules to improve large-region segmentation. Experiments on a real underwater fouling dataset show that the proposed model achieves 92.41% IoU, 96.03% DSC, and 96.14% accuracy, outperforming baseline models and state-of-the-art segmentation methods. This lightweight and robust framework supports practical intelligent inspection for marine engineering applications. Code is available at <a href="https://github.com/lin2000921/fma-unext.git">https://github.com/lin2000921/fma-unext.git</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

Frequency-adaptive and multi-scale attention network for underwater hull fouling segmentation

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Yihao Lin,
  • Yajuan Gu,
  • Lunming Qin,
  • Liang Xue,
  • Houqin Bian,
  • Xi Wang

摘要

Marine biofouling on ship hulls reduces navigation efficiency and increases maintenance costs, making accurate semantic segmentation critical for intelligent ship maintenance. Underwater imaging suffers from low contrast, uneven illumination, and large fouling scale variations, which degrade standard segmentation models. This work presents an enhanced UNeXt architecture named FMA-UNeXt for underwater hull fouling segmentation. The model integrates frequency-adaptive dilated convolution to enhance high-frequency details and low-frequency structures, a multi-scale convolutional attention decoder to handle scale variations, and efficient multi-scale attention modules to improve large-region segmentation. Experiments on a real underwater fouling dataset show that the proposed model achieves 92.41% IoU, 96.03% DSC, and 96.14% accuracy, outperforming baseline models and state-of-the-art segmentation methods. This lightweight and robust framework supports practical intelligent inspection for marine engineering applications. Code is available at https://github.com/lin2000921/fma-unext.git