Scholarlink

<p>In recent years, image super-resolution (SR) research has made remarkable progress. However, the complexity of the models, such as increased network depth, attention mechanisms, and Transformer structures, has resulted in high computational costs, making it challenging to deploy these models on mobile devices. To address this issue, we propose a lightweight SR model based on the group residual feature network (GRFN). Our model features an efficient Group Residual Feature Block (GRFB), composed mainly of Feature Pixel Convolution Units (FPUs) combined with local residual connection stacking. This design simplifies feature fusion and strikes a balance between model performance and inference time. With a scale factor of 4, the number of parameters of GRFN is only 554&#xa0;K, and PSNR/SSIM is able to reach 32.30&#xa0;dB/0.8965, which balances the inference speed and quality and outperforms the state-of-the-art methods. Furthermore, our model achieves excellent reconstruction results in subjective visual evaluations.</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

GRFN: A Group Residual Feature Network for Lightweight Image Super-Resolution

  • 期刊论文

  • 发表日期：

    2025年1月9日

查看全文
  • Xin Yang,
  • Chaming Hong,
  • Panpan Zhang

摘要

In recent years, image super-resolution (SR) research has made remarkable progress. However, the complexity of the models, such as increased network depth, attention mechanisms, and Transformer structures, has resulted in high computational costs, making it challenging to deploy these models on mobile devices. To address this issue, we propose a lightweight SR model based on the group residual feature network (GRFN). Our model features an efficient Group Residual Feature Block (GRFB), composed mainly of Feature Pixel Convolution Units (FPUs) combined with local residual connection stacking. This design simplifies feature fusion and strikes a balance between model performance and inference time. With a scale factor of 4, the number of parameters of GRFN is only 554 K, and PSNR/SSIM is able to reach 32.30 dB/0.8965, which balances the inference speed and quality and outperforms the state-of-the-art methods. Furthermore, our model achieves excellent reconstruction results in subjective visual evaluations.