Scholarlink

Aiming at the problem that image super-resolution reconstruction networks directly introduce the attention mechanism into the residual blocks while ignoring the residual features themselves and simply fusing different levels of features, a hierarchical directional attention fusion network (HDAFN) was designed. Firstly, a bifurcated attention residual block is proposed, which obtains attention features without adding additional parameters and retains the original residual features. Secondly, non-local coordinate attention is designed to reduce the computational burden of self-focused features. Finally, an orientation attention fusion module is introduced to effectively integrate the attention residual features using different orientation information. Extensive experimental results show that HDAFN performs well in quantifying the peak signal-to-noise ratio and testing structural similarity of the four benchmark datasets. Specifically, HDAFN increases the number of parameters by only 149 K and the peak signal-to-noise ratio (PSNR) by 0.23 dB on average after adding the proposed module to the residual block.

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

Hierarchical Directional Attention Fusion Network for Image Super Resolution

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • He Jiang,
  • Chentao Zhu,
  • Rui Zhang,
  • Xin Huang,
  • Zhiwei Cheng,
  • Peijie Wang,
  • Hao Yin,
  • Chang Liu,
  • Jing Yuan,
  • Haoran Zuo,
  • Junjie Li,
  • Boming Song,
  • Deqiang Cheng

摘要

Aiming at the problem that image super-resolution reconstruction networks directly introduce the attention mechanism into the residual blocks while ignoring the residual features themselves and simply fusing different levels of features, a hierarchical directional attention fusion network (HDAFN) was designed. Firstly, a bifurcated attention residual block is proposed, which obtains attention features without adding additional parameters and retains the original residual features. Secondly, non-local coordinate attention is designed to reduce the computational burden of self-focused features. Finally, an orientation attention fusion module is introduced to effectively integrate the attention residual features using different orientation information. Extensive experimental results show that HDAFN performs well in quantifying the peak signal-to-noise ratio and testing structural similarity of the four benchmark datasets. Specifically, HDAFN increases the number of parameters by only 149 K and the peak signal-to-noise ratio (PSNR) by 0.23 dB on average after adding the proposed module to the residual block.