Scholarlink

<p>The goal of super-resolution is to recover a high-resolution (HR) image from its low-resolution counterpart. To address the issues of high-frequency information loss and insufficient global context modeling in high-resolution reconstruction, this paper proposes a multi-scale fusion inference module called Prior Prompt Image Restoration Block (PPIRB), based on the “result prompt process”. We innovatively employ a prior-generated approximation of the HR image as prompt information, which guides the network in a progressive manner to enhance long-range semantic dependencies and the recovery of high-frequency details.Extensive experimental results show that the super-resolution network integrated with PPIRB module achieves good performance improvement at different magnifications of multiple benchmark datasets, especially in complex texture scenes, where the ability to preserve high-frequency details is significantly enhanced. The effectiveness of PPIRB in expanding the context aware range of the model and enhancing long-range spatial dependencies is further validated through Local Attribution Maps (LAM) and Effective Receptive Field (ERF) analysis. In addition, PPIRB is a lightweight module that allows for seamless integration into or direct replacement of reconstruction modules in existing super-resolution networks.</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

PPIRB:generate prior prompt information to dynamically adjust the high-resolution reconstruction process

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Guibao Wang,
  • Yiheng Ren,
  • Mingyuan Xue,
  • Shuzhen Wang

摘要

The goal of super-resolution is to recover a high-resolution (HR) image from its low-resolution counterpart. To address the issues of high-frequency information loss and insufficient global context modeling in high-resolution reconstruction, this paper proposes a multi-scale fusion inference module called Prior Prompt Image Restoration Block (PPIRB), based on the “result prompt process”. We innovatively employ a prior-generated approximation of the HR image as prompt information, which guides the network in a progressive manner to enhance long-range semantic dependencies and the recovery of high-frequency details.Extensive experimental results show that the super-resolution network integrated with PPIRB module achieves good performance improvement at different magnifications of multiple benchmark datasets, especially in complex texture scenes, where the ability to preserve high-frequency details is significantly enhanced. The effectiveness of PPIRB in expanding the context aware range of the model and enhancing long-range spatial dependencies is further validated through Local Attribution Maps (LAM) and Effective Receptive Field (ERF) analysis. In addition, PPIRB is a lightweight module that allows for seamless integration into or direct replacement of reconstruction modules in existing super-resolution networks.