Scholarlink

<p>Single-image super-resolution (SISR) methods incorporate attention mechanisms to emphasize important features to improve the recovery of corresponding high-resolution images. However, the features are commonly learned from scratch by the attention mechanisms without explicit guidance, thereby degrading the learning efficiency and reconstruction accuracy of the network. In this paper, we propose a simple yet effective attention mechanism guided by inherent high-frequency information of images, called High-frequency Guiding Attention (HGA), to enhance representation capability of network. Building on HGA, we introduce a High-frequency Information Guiding Attention Network (HIGAN) for SISR, in which precise high-frequency information are extracted to improve the SR results. The backbone of HIGAN consists of a sequence of Dual-Branch Frequency Refining blocks (DFRBs). Each DFRB contains a Low-Frequency Refining Block (LFRB), a High-Frequency Guiding Attention Block (HGAB), and a skip connection to handle features across different frequency bands. The HGAB employs an Initial High-frequency Extraction Block (IHEB) to capture high-frequency information from the input and then guides the subsequent High-frequency Enhancing Block (HFEB) to focus on the most informative information. Meanwhile, the LFRB produces more intermediate frequency information for SR results. Extensive experiments on benchmark datasets demonstrate that HIGAN not only outperforms state-of-the-art methods but also remains highly efficient due to its lightweight architecture. The code for our proposed method will be available at <a href="https://github.com/TyVirgo/HIGAN">https://github.com/TyVirgo/HIGAN</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

A high-frequency information guiding attention network for super-lightweight image super-resolution

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Yuanhui Tian,
  • Jinsheng Fang,
  • Yuhui Li,
  • Kun Zeng

摘要

Single-image super-resolution (SISR) methods incorporate attention mechanisms to emphasize important features to improve the recovery of corresponding high-resolution images. However, the features are commonly learned from scratch by the attention mechanisms without explicit guidance, thereby degrading the learning efficiency and reconstruction accuracy of the network. In this paper, we propose a simple yet effective attention mechanism guided by inherent high-frequency information of images, called High-frequency Guiding Attention (HGA), to enhance representation capability of network. Building on HGA, we introduce a High-frequency Information Guiding Attention Network (HIGAN) for SISR, in which precise high-frequency information are extracted to improve the SR results. The backbone of HIGAN consists of a sequence of Dual-Branch Frequency Refining blocks (DFRBs). Each DFRB contains a Low-Frequency Refining Block (LFRB), a High-Frequency Guiding Attention Block (HGAB), and a skip connection to handle features across different frequency bands. The HGAB employs an Initial High-frequency Extraction Block (IHEB) to capture high-frequency information from the input and then guides the subsequent High-frequency Enhancing Block (HFEB) to focus on the most informative information. Meanwhile, the LFRB produces more intermediate frequency information for SR results. Extensive experiments on benchmark datasets demonstrate that HIGAN not only outperforms state-of-the-art methods but also remains highly efficient due to its lightweight architecture. The code for our proposed method will be available at https://github.com/TyVirgo/HIGAN.