Scholarlink

<p>In this paper, we study reference-based image super-resolution (RefSR), where external reference images provide high-resolution cues and a higher texture compensation than single image super-resolution (SISR), but are not suited for cross domain. High-resolution images and low-resolution input often contain large color and geometric mismatches, which can produce strong artifacts directly aligned with the input. Single-path designs also do not provide local detail quality and global structural consistency, and may obtain inaccurate texture recovery or non-significant deformation. Other unreliable reference features may also be erroneous and consequently degrade reconstruction stability. We design a Structure-Guided Collaborative Network (SGCNet) consisting of three aspects. A Lightweight Color and Structure Alignment Module (LCAM) performs cross-domination registration based on color normalization and structure-consistency mask to minimize style drift and misalignment artifacts. A Dual-Path Encoder where a multiscale CNN branch and a structure aware Mamba branch are used to capture local high-frequency patterns and long-range structural dependency controls unreliable reference cues. A Structure-Difference-Aware Gated Fusion Module (SGFM) controls unreliable sources using difference-enabled gating to achieve consistent structure–texture fusion at scale. Extensive experiments have shown that SGCNet performs better than existing state-of-the-art techniques in both quantitative and qualitative metrics.</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

Structure-guided collaborative network for cross-domain reference-based remote sensing image super-resolution

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月17日

查看全文
  • Xiaojiao Tao,
  • Qianying Feng,
  • Yinghua Li,
  • Weiao Hao,
  • Yining Zhang

摘要

In this paper, we study reference-based image super-resolution (RefSR), where external reference images provide high-resolution cues and a higher texture compensation than single image super-resolution (SISR), but are not suited for cross domain. High-resolution images and low-resolution input often contain large color and geometric mismatches, which can produce strong artifacts directly aligned with the input. Single-path designs also do not provide local detail quality and global structural consistency, and may obtain inaccurate texture recovery or non-significant deformation. Other unreliable reference features may also be erroneous and consequently degrade reconstruction stability. We design a Structure-Guided Collaborative Network (SGCNet) consisting of three aspects. A Lightweight Color and Structure Alignment Module (LCAM) performs cross-domination registration based on color normalization and structure-consistency mask to minimize style drift and misalignment artifacts. A Dual-Path Encoder where a multiscale CNN branch and a structure aware Mamba branch are used to capture local high-frequency patterns and long-range structural dependency controls unreliable reference cues. A Structure-Difference-Aware Gated Fusion Module (SGFM) controls unreliable sources using difference-enabled gating to achieve consistent structure–texture fusion at scale. Extensive experiments have shown that SGCNet performs better than existing state-of-the-art techniques in both quantitative and qualitative metrics.