Scholarlink

<p>Pansharpening, the fusion of a high-resolution panchromatic (PAN) image with low-resolution multispectral (LRMS) imagery to produce high-resolution multispectral (HRMS) outputs, remains a core problem in remote sensing. Existing approaches often underemphasize the joint integration of global spatial context and local texture information, which can lead to localized noise and spectral or spatial distortions. In this work, we introduce HMTCnet, a hybrid Transformer–convolutional multi-stream network for pansharpening, which integrates a Spatial Feature Fusion Network (SFN) and a Texture Feature Fusion Network (TFN) through a structured multi-stream collaborative and adaptive fusion mechanism. The SFN employs a self-attention encoder to capture global spatial context and a fusion mechanism to generate high-resolution spatial features (HRT). Subsequently, the PAN image is used as spatial guidance information and aligned with HRT in the Transformer through attention-based cross-spectral feature alignment. The TFN adopts a multi-stream encoder–decoder design to extract, upsample, and compress texture representations. SFN and TFN are connected to enable efficient multi-scale fusion of spatial and texture features, producing the final HRMS result. Extensive experiments on two widely used benchmark datasets demonstrate that HMTCnet significantly outperforms state-of-the-art methods on standard image-quality metrics. The code is publicly available and can be accessed at <a href="https://github.com/jiaruihsi2025/HMTCnet">https://github.com/jiaruihsi2025/HMTCnet</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 multi-stream network for pansharpening based on transformer and convolution

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Jiarui Shi,
  • Yue Pan,
  • Zhongtao Li,
  • Shanshan Wang,
  • Zilong Zhao,
  • Minglei Wang,
  • Zheng Cao

摘要

Pansharpening, the fusion of a high-resolution panchromatic (PAN) image with low-resolution multispectral (LRMS) imagery to produce high-resolution multispectral (HRMS) outputs, remains a core problem in remote sensing. Existing approaches often underemphasize the joint integration of global spatial context and local texture information, which can lead to localized noise and spectral or spatial distortions. In this work, we introduce HMTCnet, a hybrid Transformer–convolutional multi-stream network for pansharpening, which integrates a Spatial Feature Fusion Network (SFN) and a Texture Feature Fusion Network (TFN) through a structured multi-stream collaborative and adaptive fusion mechanism. The SFN employs a self-attention encoder to capture global spatial context and a fusion mechanism to generate high-resolution spatial features (HRT). Subsequently, the PAN image is used as spatial guidance information and aligned with HRT in the Transformer through attention-based cross-spectral feature alignment. The TFN adopts a multi-stream encoder–decoder design to extract, upsample, and compress texture representations. SFN and TFN are connected to enable efficient multi-scale fusion of spatial and texture features, producing the final HRMS result. Extensive experiments on two widely used benchmark datasets demonstrate that HMTCnet significantly outperforms state-of-the-art methods on standard image-quality metrics. The code is publicly available and can be accessed at https://github.com/jiaruihsi2025/HMTCnet.