Scholarlink

<p>Effectively leveraging both high- and low-frequency cues from low-resolution images remains a key challenge in single image super-resolution(SISR). To address this, we propose the Local–global synergistic Enhanced Perception Network (LSEPN), a lightweight CNN-Transformer cooperative model: The local branch employs multi-type gradient-aware differential convolutions to explicitly encode edge and texture priors, significantly enhancing the detail representation and generalization of vanilla convolutions; the global branch captures long-range dependencies and low-frequency semantics via multi-level spatial self-attention (local, cross-region, global). Additionally, we introduce a Frequency-Aware Fusion Module(FAFM) that adaptively balances structural preservation and detail enhancement for feature fusion. Visualizations of Fourier spectrum and Local attribution map reveal reduced feature redundancy and broader pixel activation. LSEPN achieves competitive gains (e.g., +0.49dB on Manga109 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>4) over state-of-the-art hybrids (e.g., HCFormer) while simultaneously reducing parameters (26.2%) and computation (19.4%), facilitating efficient inference and scalable parallel processing on modern GPUs.</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

Local–global synergistic enhanced perception network for lightweight super-resolution

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Liang Dong,
  • Haicheng Zhang,
  • Guantao Wang,
  • Guiling Sun

摘要

Effectively leveraging both high- and low-frequency cues from low-resolution images remains a key challenge in single image super-resolution(SISR). To address this, we propose the Local–global synergistic Enhanced Perception Network (LSEPN), a lightweight CNN-Transformer cooperative model: The local branch employs multi-type gradient-aware differential convolutions to explicitly encode edge and texture priors, significantly enhancing the detail representation and generalization of vanilla convolutions; the global branch captures long-range dependencies and low-frequency semantics via multi-level spatial self-attention (local, cross-region, global). Additionally, we introduce a Frequency-Aware Fusion Module(FAFM) that adaptively balances structural preservation and detail enhancement for feature fusion. Visualizations of Fourier spectrum and Local attribution map reveal reduced feature redundancy and broader pixel activation. LSEPN achieves competitive gains (e.g., +0.49dB on Manga109 \(\times \) × 4) over state-of-the-art hybrids (e.g., HCFormer) while simultaneously reducing parameters (26.2%) and computation (19.4%), facilitating efficient inference and scalable parallel processing on modern GPUs.