Scholarlink

<p>Recent transformer-based architectures have achieved remarkable progress in single-image super-resolution, owing largely to the strength of window-based self-attention (WSA) in capturing local dependencies. However, the fixed receptive fields of WSA constrain its ability to effectively model global context, and enlarging them leads to excessive computational cost. Meanwhile, existing deformable attention mechanisms, though efficient in high-level vision tasks, often fail to preserve the fine-grained spatial details essential for visually plausible reconstruction in low-level graphics problems. To overcome these challenges, we present DWGLT, a Deformable WSA-based Global–Local Transformer network specifically designed for high-fidelity image reconstruction. DWGLT integrates deformable receptive field learning into Global–Local Transformer blocks that jointly capture global semantics and fine details. The global feature extraction block leverages a deformable receptive field aggregation module to adaptively gather long-range information and a global-guided cross attention module to refine local features under global context. The local feature extraction block further enhances detail recovery through overlapping receptive fields, while a spatial-channel compression transformation reduces computational overhead without degrading perceptual quality. Extensive experiments on multiple benchmarks demonstrate that DWGLT achieves state-of-the-art reconstruction accuracy and visual realism with high efficiency, making it a promising solution for image restoration and visual content enhancement in computer graphics applications. The code is available at <a href="https://github.com/CCGO2t2/DWGLT">https://github.com/CCGO2t2/DWGLT</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

DWGLT: a deformable window-based global–local transformer network for efficient image super-resolution

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Jian Zhu,
  • Simin Zheng,
  • Peiyao Chen,
  • Jiale Zhao,
  • Zhaohui Liao,
  • Rongqiu Wang,
  • Guang Feng

摘要

Recent transformer-based architectures have achieved remarkable progress in single-image super-resolution, owing largely to the strength of window-based self-attention (WSA) in capturing local dependencies. However, the fixed receptive fields of WSA constrain its ability to effectively model global context, and enlarging them leads to excessive computational cost. Meanwhile, existing deformable attention mechanisms, though efficient in high-level vision tasks, often fail to preserve the fine-grained spatial details essential for visually plausible reconstruction in low-level graphics problems. To overcome these challenges, we present DWGLT, a Deformable WSA-based Global–Local Transformer network specifically designed for high-fidelity image reconstruction. DWGLT integrates deformable receptive field learning into Global–Local Transformer blocks that jointly capture global semantics and fine details. The global feature extraction block leverages a deformable receptive field aggregation module to adaptively gather long-range information and a global-guided cross attention module to refine local features under global context. The local feature extraction block further enhances detail recovery through overlapping receptive fields, while a spatial-channel compression transformation reduces computational overhead without degrading perceptual quality. Extensive experiments on multiple benchmarks demonstrate that DWGLT achieves state-of-the-art reconstruction accuracy and visual realism with high efficiency, making it a promising solution for image restoration and visual content enhancement in computer graphics applications. The code is available at https://github.com/CCGO2t2/DWGLT.