Scholarlink

<p>In recent years, attention mechanisms have achieved significant results in image super-resolution tasks, effectively improving reconstruction quality by enhancing the focus on key features. However, existing methods still have limitations, such as the difficulty in simultaneously modeling long-range dependencies and local details with either a single spatial-wise or channel-wise self-attention mechanism, resulting in an imbalance between global structure recovery and high-frequency information detail restoration. To address this, we propose a lightweight global information aggregation network (GIAN). Specifically, our proposed global cross self-attention module adaptively adjusts channel weights and captures spatial relationships between pixels by coupling channel-wise and spatial-wise self-attention mechanisms, significantly improving local detail restoration. Additionally, in order to achieve efficient fusion of hierarchical features, we design an aggregation-gated feedforward network module that utilizes depth-wise convolution and channel concatenation operations while enhancing the representation of key features through gating mechanisms. Experiments demonstrate that, compared to state-of-the-art lightweight super-resolution methods, GIAN achieves superior performance on multiple benchmark datasets. Code is available at <a href="https://github.com/SJHunag/GIAN">https://github.com/SJHunag/GIAN</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

Global information aggregation network for lightweight single image super-resolution

  • 期刊论文

  • 发表日期：

    2025年7月10日

查看全文
  • Shijun Huang,
  • Shigang Hu,
  • Guoyi Zheng,
  • Jianxin Wang

摘要

In recent years, attention mechanisms have achieved significant results in image super-resolution tasks, effectively improving reconstruction quality by enhancing the focus on key features. However, existing methods still have limitations, such as the difficulty in simultaneously modeling long-range dependencies and local details with either a single spatial-wise or channel-wise self-attention mechanism, resulting in an imbalance between global structure recovery and high-frequency information detail restoration. To address this, we propose a lightweight global information aggregation network (GIAN). Specifically, our proposed global cross self-attention module adaptively adjusts channel weights and captures spatial relationships between pixels by coupling channel-wise and spatial-wise self-attention mechanisms, significantly improving local detail restoration. Additionally, in order to achieve efficient fusion of hierarchical features, we design an aggregation-gated feedforward network module that utilizes depth-wise convolution and channel concatenation operations while enhancing the representation of key features through gating mechanisms. Experiments demonstrate that, compared to state-of-the-art lightweight super-resolution methods, GIAN achieves superior performance on multiple benchmark datasets. Code is available at https://github.com/SJHunag/GIAN.