Scholarlink

<p>In recent years, convolutional neural networks (CNNs) and Transformers have been widely explored for face image super-resolution tasks. However, CNNs tend to compromise the fidelity and naturalness of reconstructed images due to their emphasis on local information extraction, while Transformers’ quadratic complexity limits their application to low-level vision tasks. To address these challenges, we introduce Visual Mamba, a state-space model that exhibits linear complexity and a global receptive field, for face super-resolution. We propose a CNN-Mamba Cooperation Network (CMCNet) that utilizes CNN to capture local facial dependencies and Mamba to model global facial structures. To mitigate the pixel forgetting problem of Mamba in low-level vision tasks, we introduce a multi-scale 2D scanning module (MS2D). Additionally, a dual-branch cross-fusion module (DCFM) is designed to effectively integrate local and global information. A multi-scale feature fusion module (MFFM) is also employed to adaptively fuse features from different stages of the encoder. Extensive experiments on benchmark datasets demonstrate that CMCNet consistently outperforms state-of-the-art methods, highlighting its potential for high-quality face super-resolution. Code is available at <a href="https://github.com/axhryy/CMCNet">https://github.com/axhryy/CMCNet</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

CMCNet: enhancing face image super-resolution through CNN-Mamba collaboration

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Zhaoci Liu,
  • Huayong Liu

摘要

In recent years, convolutional neural networks (CNNs) and Transformers have been widely explored for face image super-resolution tasks. However, CNNs tend to compromise the fidelity and naturalness of reconstructed images due to their emphasis on local information extraction, while Transformers’ quadratic complexity limits their application to low-level vision tasks. To address these challenges, we introduce Visual Mamba, a state-space model that exhibits linear complexity and a global receptive field, for face super-resolution. We propose a CNN-Mamba Cooperation Network (CMCNet) that utilizes CNN to capture local facial dependencies and Mamba to model global facial structures. To mitigate the pixel forgetting problem of Mamba in low-level vision tasks, we introduce a multi-scale 2D scanning module (MS2D). Additionally, a dual-branch cross-fusion module (DCFM) is designed to effectively integrate local and global information. A multi-scale feature fusion module (MFFM) is also employed to adaptively fuse features from different stages of the encoder. Extensive experiments on benchmark datasets demonstrate that CMCNet consistently outperforms state-of-the-art methods, highlighting its potential for high-quality face super-resolution. Code is available at https://github.com/axhryy/CMCNet.