Scholarlink

<p>Lightweight image super-resolution (SR) methods prioritize computational efficiency but often rely on restricted receptive fields or shallow architectures, making it difficult to recover complex textures and long-range structures. Furthermore, conventional attention mechanisms are typically confined to spatial or channel domains, underutilizing frequency information and leading to significant high-frequency detail loss. To address these limitations, we propose the Lightweight Frequency-Spatial Distillation Attention Network (FSDAN), aimed at coordinating global structural consistency and local high-frequency textures under strict parameter constraints. Specifically, we construct the Frequency-Spatial Interaction Unit (FSIU), which employs Fast Fourier Convolution (FFC) to capture global frequency structures and multi-scale Blueprint Separable Convolutions (BSConv) to extract hierarchical spatial textures. This dual-branch mechanism effectively enhances both long-range contextual and local detail representations. Building upon this, the Progressive Frequency-Spatial Fusion Block (PFSFB) facilitates layer-by-layer feature enhancement-from local textures to global structures-through multi-level distillation and phased fusion. Additionally, the Frequency-Domain Guided Attention Enhancement Module (FGAEM) utilizes global frequency priors to dynamically regulate multi-scale spatial attention, fostering robust cross-scale interaction and texture refinement. Extensive experiments demonstrate that FSDAN achieves highly competitive performance against state-of-the-art lightweight models, offering a favorable trade-off among complexity, reconstruction quality, and inference speed. The code and pre-trained models are available at <a href="https://github.com/Anytime23/FSDAN">https://github.com/Anytime23/FSDAN</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 frequency-spatial distillation attention network for image super-resolution

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Siwei Wei,
  • Tianwei Xu,
  • Xueli Chang,
  • Chunzhi Wang

摘要

Lightweight image super-resolution (SR) methods prioritize computational efficiency but often rely on restricted receptive fields or shallow architectures, making it difficult to recover complex textures and long-range structures. Furthermore, conventional attention mechanisms are typically confined to spatial or channel domains, underutilizing frequency information and leading to significant high-frequency detail loss. To address these limitations, we propose the Lightweight Frequency-Spatial Distillation Attention Network (FSDAN), aimed at coordinating global structural consistency and local high-frequency textures under strict parameter constraints. Specifically, we construct the Frequency-Spatial Interaction Unit (FSIU), which employs Fast Fourier Convolution (FFC) to capture global frequency structures and multi-scale Blueprint Separable Convolutions (BSConv) to extract hierarchical spatial textures. This dual-branch mechanism effectively enhances both long-range contextual and local detail representations. Building upon this, the Progressive Frequency-Spatial Fusion Block (PFSFB) facilitates layer-by-layer feature enhancement-from local textures to global structures-through multi-level distillation and phased fusion. Additionally, the Frequency-Domain Guided Attention Enhancement Module (FGAEM) utilizes global frequency priors to dynamically regulate multi-scale spatial attention, fostering robust cross-scale interaction and texture refinement. Extensive experiments demonstrate that FSDAN achieves highly competitive performance against state-of-the-art lightweight models, offering a favorable trade-off among complexity, reconstruction quality, and inference speed. The code and pre-trained models are available at https://github.com/Anytime23/FSDAN.