Scholarlink

<p>Transformer-based super-resolution(SR) reconstruction methods have shown significant performance because transformer’s self-attention (SA) mechanism explores non-local information to obtain better high-resolution(HR) image reconstruction. However, the low-pass filtering nature of the SA mechanism limits its ability to capture local details, leading to overly smooth reconstruction results. Furthermore, the transformer is less effective at constructing high-frequency representations compared to conventional convolutional structures. To address these issues, we propose a Global Self-Modulating Feature Aggregation Module (GSMFAM) to exploit the interaction between local and global features for more accurate reconstruction. Specifically, GSMFAM uses High-Frequency Feature Extraction (HFFE) branch to capture local information and Efficient Self-Modulation Characteristics (ESMC) branch to capture global information. In addition, we designed a Large Kernel-Guided Attention Module(LKGAM) to optimize the feature fusion process and achieve better reconstruction results. Extensive experiments have shown that the proposed self-modulated feature aggregation network (SMANet) has undergone extensive ablation studies and experiments (e.g., Carbonate2D and Sandstone2D) in SR tasks to evaluate its performance. The code is available at <a href="https://github.com/ChaoHan3/SMANet/tree/master">https://github.com/ChaoHan3/SMANet/tree/master</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

SMANET:self-modulating feature aggregation networks for image super-resolution in digital cores

  • 期刊论文

  • 发表日期：

    2025年9月13日

查看全文
  • Yubo Zhang,
  • Chao Han,
  • Junhao Bi,
  • Lei Xu,
  • Haibin Xiang,
  • Haihua Kong

摘要

Transformer-based super-resolution(SR) reconstruction methods have shown significant performance because transformer’s self-attention (SA) mechanism explores non-local information to obtain better high-resolution(HR) image reconstruction. However, the low-pass filtering nature of the SA mechanism limits its ability to capture local details, leading to overly smooth reconstruction results. Furthermore, the transformer is less effective at constructing high-frequency representations compared to conventional convolutional structures. To address these issues, we propose a Global Self-Modulating Feature Aggregation Module (GSMFAM) to exploit the interaction between local and global features for more accurate reconstruction. Specifically, GSMFAM uses High-Frequency Feature Extraction (HFFE) branch to capture local information and Efficient Self-Modulation Characteristics (ESMC) branch to capture global information. In addition, we designed a Large Kernel-Guided Attention Module(LKGAM) to optimize the feature fusion process and achieve better reconstruction results. Extensive experiments have shown that the proposed self-modulated feature aggregation network (SMANet) has undergone extensive ablation studies and experiments (e.g., Carbonate2D and Sandstone2D) in SR tasks to evaluate its performance. The code is available at https://github.com/ChaoHan3/SMANet/tree/master.