Scholarlink

<p>Deep learning has significantly advanced single-image super-resolution (SISR) tasks, although it often demands high computational resources. Decreasing computational complexity in deep learning involves reducing network parameters. Models with fewer parameters not only lead to faster inference times and reduced memory requirements but also enhance generalization capabilities. In this paper, we propose an innovative approach to reduce parameters in SISR models without compromising performance. By splitting input channels, processing them independently, and then recombining them, we reduced the number of model parameters for learning while maintaining input dimensionality. Our method contributes to developing resource-efficient SISR models that strike a balance between accuracy and computational demands. We evaluated the proposed method on five benchmark datasets across three scaling factors. The experimental results demonstrate that our approach reduces the number of parameters, Memory usages and floating-point operations (FLOPs) by 47.5% considerably without significant reduction of quality. Therefore, this approach achieves an acceptable trade-off between the number of model parameters and result accuracy. Our code Implementation can be found at <a href="https://github.com/AminTolou/RC-SPAN">https://github.com/AminTolou/RC-SPAN</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

RC-SPAN: a low-complexity RCAN by channel splitting module for single-image super-resolution

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Mohammad Amin Tolou Beydokhti,
  • Alireza Ahmadyfard,
  • Hossein Khosravi

摘要

Deep learning has significantly advanced single-image super-resolution (SISR) tasks, although it often demands high computational resources. Decreasing computational complexity in deep learning involves reducing network parameters. Models with fewer parameters not only lead to faster inference times and reduced memory requirements but also enhance generalization capabilities. In this paper, we propose an innovative approach to reduce parameters in SISR models without compromising performance. By splitting input channels, processing them independently, and then recombining them, we reduced the number of model parameters for learning while maintaining input dimensionality. Our method contributes to developing resource-efficient SISR models that strike a balance between accuracy and computational demands. We evaluated the proposed method on five benchmark datasets across three scaling factors. The experimental results demonstrate that our approach reduces the number of parameters, Memory usages and floating-point operations (FLOPs) by 47.5% considerably without significant reduction of quality. Therefore, this approach achieves an acceptable trade-off between the number of model parameters and result accuracy. Our code Implementation can be found at https://github.com/AminTolou/RC-SPAN.