Scholarlink

<p>Accurate semantic segmentation of underwater images, such as those depicting fish, coral reefs, and shipwrecks, is pivotal for marine ecological monitoring and resource exploration. While convolutional neural networks (CNNs) excel in extracting local information, their limited receptive field hinders the understanding of context dependency. To address this, we introduce Z-Net, a hybrid network that integrates Transformer and CNN branches. The Transformer branch captures global context from low-resolution inputs, reducing computational costs, while the CNN branch extracts fine local details. We propose a feature restoration module (FRM) to recover high-resolution features from low-resolution ones, enhancing detailed feature refinement. Additionally, a multi-attention fusion module (MAFM) fuses multi-scale features from both branches, further improving segmentation accuracy. Comprehensive experiments on the SUIM dataset and real-world underwater images demonstrate that Z-Net significantly outperforms existing methods, highlighting its effectiveness for underwater image semantic segmentation. The code is available at <a href="https://github.com/zhangjx0613/ZNet">https://github.com/zhangjx0613/ZNet</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

Hybrid transformer-CNN architecture for enhanced underwater image semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Jiaxing Zhang,
  • Yujuan Sun,
  • Hua Wang,
  • Xiaofeng Zhang

摘要

Accurate semantic segmentation of underwater images, such as those depicting fish, coral reefs, and shipwrecks, is pivotal for marine ecological monitoring and resource exploration. While convolutional neural networks (CNNs) excel in extracting local information, their limited receptive field hinders the understanding of context dependency. To address this, we introduce Z-Net, a hybrid network that integrates Transformer and CNN branches. The Transformer branch captures global context from low-resolution inputs, reducing computational costs, while the CNN branch extracts fine local details. We propose a feature restoration module (FRM) to recover high-resolution features from low-resolution ones, enhancing detailed feature refinement. Additionally, a multi-attention fusion module (MAFM) fuses multi-scale features from both branches, further improving segmentation accuracy. Comprehensive experiments on the SUIM dataset and real-world underwater images demonstrate that Z-Net significantly outperforms existing methods, highlighting its effectiveness for underwater image semantic segmentation. The code is available at https://github.com/zhangjx0613/ZNet.