Scholarlink

<p>Single Image Super-Resolution (SISR) based on convolutional neural networks (CNNs) has achieved significant progress. However, current single-branch SR architectures struggle to simultaneously restore low and high-frequency information from low-resolution (LR) images. In this paper, we propose a Cross-Scale Atomic Feature Enhanced Network (CSAFEN) for SISR. CSAFEN consists of a Reconstruction Subnetwork (RS) and a High-Frequency Compensation Subnetwork (HFCS). A Cross-Scale High-Frequency Extraction Module (CSHE) is introduced to exploit inherent high-frequency priors from LR images to compensate for reconstructed textures and details. Furthermore, an Atomic-Features Matching Optimizer (AMO) is developed to overcome the limitations of convolutional features, which are prone to losing details. AMO constructs a high-resolution dictionary and matches input features with cross-scale atomic features, selecting relevant high-resolution features to compensate for the LR input. Extensive experiments demonstrate that CSAFEN outperforms state-of-the-art SISR methods in terms of both quantitative metrics and visual quality, especially for images rich in structural and textural information.</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

Cross-Scale Atomic Feature Enhanced Network for high-fidelity Single Image Super-Resolution

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Aiping Yang,
  • Chenhui Yu,
  • Jinbin Wang,
  • Zihao Wei,
  • Jiale Cao,
  • Liping Liu

摘要

Single Image Super-Resolution (SISR) based on convolutional neural networks (CNNs) has achieved significant progress. However, current single-branch SR architectures struggle to simultaneously restore low and high-frequency information from low-resolution (LR) images. In this paper, we propose a Cross-Scale Atomic Feature Enhanced Network (CSAFEN) for SISR. CSAFEN consists of a Reconstruction Subnetwork (RS) and a High-Frequency Compensation Subnetwork (HFCS). A Cross-Scale High-Frequency Extraction Module (CSHE) is introduced to exploit inherent high-frequency priors from LR images to compensate for reconstructed textures and details. Furthermore, an Atomic-Features Matching Optimizer (AMO) is developed to overcome the limitations of convolutional features, which are prone to losing details. AMO constructs a high-resolution dictionary and matches input features with cross-scale atomic features, selecting relevant high-resolution features to compensate for the LR input. Extensive experiments demonstrate that CSAFEN outperforms state-of-the-art SISR methods in terms of both quantitative metrics and visual quality, especially for images rich in structural and textural information.