<p>Recently, there has been a growing interest in developing lightweight models for single-image super-resolution (SISR) in the graphics community. However, a common limitation of most existing methods is they have a restricted receptive field and struggle to utilize local–global feature information for restoration. To address this issue, we propose a lightweight SR network named LLGTN, in which a local transformer module (LTM) and a global transformer module (GTM) with adaptive attention windows are developed for efficiently extracting local–global feature information. Specifically, LTM calculates self-attention (SA) on multiple small window sizes to extract multi-scale features, which enhances the model’s capability to capture contextual information. In contrast, GTM calculates SA on a large window size by utilizing linear attention to efficiently extract global features. Extensive experimental evaluations demonstrate that LLGTN achieves state-of-the-art performance across five representative natural images SR datasets with fewer parameters and lower computational costs compared to existing lightweight SISR methods (e.g., the PSNR on the Urban100 dataset with a scaling factor of <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_3891_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="24" /> </InlineMediaObject> <EquationSource Format="TEX">\(\times 4\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>×</mo> <mn>4</mn> </mrow> </math></EquationSource> </InlineEquation> is 26.58dB with only 659K parameters). The code is available at <a href="https://github.com/Anothersongs/LLGTN">https://github.com/Anothersongs/LLGTN</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

Lightweight image super-resolution via an efficient local–global transformer network with adaptive attention windows

  • Jian Zhu,
  • Simin Zheng,
  • Peiyao Chen,
  • Min Li,
  • Jianfang Hu,
  • Ruichu Cai

摘要

Recently, there has been a growing interest in developing lightweight models for single-image super-resolution (SISR) in the graphics community. However, a common limitation of most existing methods is they have a restricted receptive field and struggle to utilize local–global feature information for restoration. To address this issue, we propose a lightweight SR network named LLGTN, in which a local transformer module (LTM) and a global transformer module (GTM) with adaptive attention windows are developed for efficiently extracting local–global feature information. Specifically, LTM calculates self-attention (SA) on multiple small window sizes to extract multi-scale features, which enhances the model’s capability to capture contextual information. In contrast, GTM calculates SA on a large window size by utilizing linear attention to efficiently extract global features. Extensive experimental evaluations demonstrate that LLGTN achieves state-of-the-art performance across five representative natural images SR datasets with fewer parameters and lower computational costs compared to existing lightweight SISR methods (e.g., the PSNR on the Urban100 dataset with a scaling factor of \(\times 4\) × 4 is 26.58dB with only 659K parameters). The code is available at https://github.com/Anothersongs/LLGTN