Scholarlink

<p>Image super-resolution (SR) is a fundamental problem in image processing. Recently, CNN-based methods have been favored for their ability to effectively utilize local information. Later, Transformer-based methods have gained significant attention for their excellent modeling of long-range pixel correlations. However, many Transformer-based methods do not simultaneously consider spatial and channel correlations, leading to incomplete feature representation. Some initial methods attempt to combine CNNs and Transformers to integrate local and global information but often fail to balance fine-grained details and long-range dependencies, resulting in suboptimal feature aggregation. To address this issue, we design a novel hybrid SR network that achieves comprehensive fusion of global and local information. We propose a Dual-path Collaborative Block (DCB), which includes a Dual Attention Transformer Module (DATM), a Multi-scale Convolution Module (MCM), and a Selective Fusion Module (SFM). Our DATM integrates channel attention with the traditional Transformer structure, enabling features to capture a broader range of information. The MCM enhances feature representation. SFM selectively fuses features based on similarity between those generated by DATM and MCM, effectively preserving the global outline and local details. Extensive experimental results on standard benchmark datasets demonstrate that our model performs better compared to existing SR methods, providing superior image quality and preservation of detail. Our code can be found at <a href="https://github.com/CZhangIR/DCNet.git">https://github.com/CZhangIR/DCNet.git</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

A CNN-transformer hybrid network with selective fusion and dual attention for image super-resolution

  • 期刊论文

  • 发表日期：

    2025年2月23日

查看全文
  • Chun Zhang,
  • Jin Wang,
  • Yunhui Shi,
  • Baocai Yin,
  • Nam Ling

摘要

Image super-resolution (SR) is a fundamental problem in image processing. Recently, CNN-based methods have been favored for their ability to effectively utilize local information. Later, Transformer-based methods have gained significant attention for their excellent modeling of long-range pixel correlations. However, many Transformer-based methods do not simultaneously consider spatial and channel correlations, leading to incomplete feature representation. Some initial methods attempt to combine CNNs and Transformers to integrate local and global information but often fail to balance fine-grained details and long-range dependencies, resulting in suboptimal feature aggregation. To address this issue, we design a novel hybrid SR network that achieves comprehensive fusion of global and local information. We propose a Dual-path Collaborative Block (DCB), which includes a Dual Attention Transformer Module (DATM), a Multi-scale Convolution Module (MCM), and a Selective Fusion Module (SFM). Our DATM integrates channel attention with the traditional Transformer structure, enabling features to capture a broader range of information. The MCM enhances feature representation. SFM selectively fuses features based on similarity between those generated by DATM and MCM, effectively preserving the global outline and local details. Extensive experimental results on standard benchmark datasets demonstrate that our model performs better compared to existing SR methods, providing superior image quality and preservation of detail. Our code can be found at https://github.com/CZhangIR/DCNet.git.