Scholarlink

<p>Blind super-resolution (BSR) targets high-fidelity reconstruction from low-resolution inputs affected by unknown and often compound degradations. Although Transformer-based restorers are effective at capturing long-range dependencies, the quadratic computational cost of standard self-attention hinders their deployment under strict resource budgets.To address this, we propose <b>LETBSR</b>, a lightweight and efficient Transformer framework for BSR that enables degradation-aware global modeling with <i>linear</i> computational complexity. Specifically, an <b>Adaptive Prior Generation Module (APGM)</b> leverages gated linear attention to infer global degradation priors with a complexity that increases linearly with respect to the sequence length <i>N</i>. Meanwhile, a <b>Content- and Degradation-driven Feature Refinement Module (CDFRM)</b> injects complementary content and degradation cues via dynamic gating, recalibrating features for enhanced detail recovery. In addition, a <b>cross-dimension collaborative CNN–Transformer backbone</b> couples global context modeling with local inductive biases, improving representational capacity without significantly inflating computational overhead.Extensive experiments on standard BSR benchmarks demonstrate that LETBSR achieves highly competitive restoration quality while substantially reducing computational costs. Compared to the state-of-the-art CDFormer, LETBSR reduces parameters by <b>42%</b> and FLOPs by <b>44%</b>, while simultaneously improving PSNR by up to <b>0.2 dB</b> on the challenging Urban100 dataset. On the same benchmark, it further surpasses DSAT by up to <b>0.5 dB</b>. The source codes are publicly available at <a href="https://github.com/helloworld7784/LETBSR">https://github.com/helloworld7784/LETBSR</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

LETBSR: Lightweight Efficient Transformer for Blind Image Super-Resolution

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Yiming Gong,
  • Yang Xu,
  • Xihuan Fang,
  • Xiangman Wu,
  • Shihang Qin

摘要

Blind super-resolution (BSR) targets high-fidelity reconstruction from low-resolution inputs affected by unknown and often compound degradations. Although Transformer-based restorers are effective at capturing long-range dependencies, the quadratic computational cost of standard self-attention hinders their deployment under strict resource budgets.To address this, we propose LETBSR, a lightweight and efficient Transformer framework for BSR that enables degradation-aware global modeling with linear computational complexity. Specifically, an Adaptive Prior Generation Module (APGM) leverages gated linear attention to infer global degradation priors with a complexity that increases linearly with respect to the sequence length N. Meanwhile, a Content- and Degradation-driven Feature Refinement Module (CDFRM) injects complementary content and degradation cues via dynamic gating, recalibrating features for enhanced detail recovery. In addition, a cross-dimension collaborative CNN–Transformer backbone couples global context modeling with local inductive biases, improving representational capacity without significantly inflating computational overhead.Extensive experiments on standard BSR benchmarks demonstrate that LETBSR achieves highly competitive restoration quality while substantially reducing computational costs. Compared to the state-of-the-art CDFormer, LETBSR reduces parameters by 42% and FLOPs by 44%, while simultaneously improving PSNR by up to 0.2 dB on the challenging Urban100 dataset. On the same benchmark, it further surpasses DSAT by up to 0.5 dB. The source codes are publicly available at https://github.com/helloworld7784/LETBSR.