Scholarlink

<p>Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs, with applications in satellite imaging, medical diagnostics, and mobile media. Recent advancements in deep learning, particularly Transformer-based models, have shown promise in capturing long-range dependencies and fine-detail reconstruction. However, challenges remain in modeling frequency-domain features and adapting to varied image content. To resolve these challenges, we design EFMSR, a Transformer-based framework that integrates Fourier Dynamic Window Attention (FDWA) and Mixture-of-Experts Pre-Attention (MEPA) modules. FDWA enhances frequency-domain detail reconstruction, while MEPA improves feature diversity and robustness. These two modules collaborate within a residual connection structure, achieving superior reconstruction performance while maintaining controllable model parameters and computational cost. Our experiments demonstrate that EFMSR outperforms state-of-the-art methods on multiple benchmarks, achieving a 0.75 dB PSNR improvement over the lightweight SwinIR baseline on the Manga109 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>4 dataset in its lightweight configuration. The code is available at <a href="https://github.com/MnTeriri/EFMSR">https://github.com/MnTeriri/EFMSR</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

Enhanced fourier-mixture transformer for high-performance image super-resolution

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Siwei Wei,
  • Shengyang Lan,
  • Feifei Wei,
  • Chunzhi Wang

摘要

Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs, with applications in satellite imaging, medical diagnostics, and mobile media. Recent advancements in deep learning, particularly Transformer-based models, have shown promise in capturing long-range dependencies and fine-detail reconstruction. However, challenges remain in modeling frequency-domain features and adapting to varied image content. To resolve these challenges, we design EFMSR, a Transformer-based framework that integrates Fourier Dynamic Window Attention (FDWA) and Mixture-of-Experts Pre-Attention (MEPA) modules. FDWA enhances frequency-domain detail reconstruction, while MEPA improves feature diversity and robustness. These two modules collaborate within a residual connection structure, achieving superior reconstruction performance while maintaining controllable model parameters and computational cost. Our experiments demonstrate that EFMSR outperforms state-of-the-art methods on multiple benchmarks, achieving a 0.75 dB PSNR improvement over the lightweight SwinIR baseline on the Manga109 \(\times \) × 4 dataset in its lightweight configuration. The code is available at https://github.com/MnTeriri/EFMSR.