Scholarlink

<p>High-resolution visual understanding demands models that balance fine local details, long-range context, and computational efficiency. We present CMT-UNet, a CNN-Mamba-Transformer hybrid architecture built on an architecture-level principle for efficient global–local coupling in dense prediction. The framework integrates a Cross-view State Space (CVSS) encoder for linear-complexity global modeling, a Region-aware Convolution Module (RCM) for spatially selective cross-branch fusion, and a Hybrid Attentional Convolution (HAC) decoder for parallel attention-convolution aggregation. Evaluated on UAVid, Vaihingen, Potsdam, and LoveDA, CMT-UNet achieves favorable accuracy–efficiency trade-offs against mainstream CNN-, Transformer-, and Mamba-based alternatives. Specifically, it reaches 69.1% mIoU on UAVid and yields evident gains for small-object segmentation, while enabling real-time inference on server-grade GPUs. Code and pre-trained models are available at <a href="https://github.com/XUE-YYC/CMT-UNet">https://github.com/XUE-YYC/CMT-UNet</a>, Zenodo DOI: <a href="https://doi.org/10.5281/zenodo.15678911">https://doi.org/10.5281/zenodo.15678911</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

CMT-UNet: enhancing remote sensing image segmentation via a hybrid CNN-Mamba-transformer architecture

  • 期刊论文

  • 发表日期：

    2026年7月19日

查看全文
  • Kang Sun,
  • Hui Chen,
  • Haoyang Zhang

摘要

High-resolution visual understanding demands models that balance fine local details, long-range context, and computational efficiency. We present CMT-UNet, a CNN-Mamba-Transformer hybrid architecture built on an architecture-level principle for efficient global–local coupling in dense prediction. The framework integrates a Cross-view State Space (CVSS) encoder for linear-complexity global modeling, a Region-aware Convolution Module (RCM) for spatially selective cross-branch fusion, and a Hybrid Attentional Convolution (HAC) decoder for parallel attention-convolution aggregation. Evaluated on UAVid, Vaihingen, Potsdam, and LoveDA, CMT-UNet achieves favorable accuracy–efficiency trade-offs against mainstream CNN-, Transformer-, and Mamba-based alternatives. Specifically, it reaches 69.1% mIoU on UAVid and yields evident gains for small-object segmentation, while enabling real-time inference on server-grade GPUs. Code and pre-trained models are available at https://github.com/XUE-YYC/CMT-UNet, Zenodo DOI: https://doi.org/10.5281/zenodo.15678911.