Scholarlink

<p>To address the challenges of high parameter counts and computational overhead in Swin Transformer-based image super-resolution methods, which hinder practical deployment, this paper proposes a lightweight dynamic adaptive fusion mixer network (DAFMixerSR). First, to enhance resource allocation flexibility, we devise a dynamic adaptive perception mixer module that combines a global predictor with streamlined contextual convolutions, effectively reducing computational complexity. Furthermore, an enhanced spatial-channel attentive (SCEA) mechanism is developed to fully integrate contextual information through spatial and channel attention, improving the efficiency of global-local feature extraction while significantly reducing parameters via depthwise separable convolutions. Additionally, a pixel-fusion convolutional feed-forward network (PFCFN) is introduced prior to the reconstruction module to further optimize restoration performance. Experimental results demonstrate that DAFMixerSR achieves an optimal balance between complexity and performance. For ×4 upscaling on the Set5 dataset, DAFMixerSR achieves PSNR improvements of 1.87 dB and 0.07 dB compared to classical SRCNN and state-of-the-art HSRNet, respectively, while requiring 46% fewer parameters than HSRNet.</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

DAFMixerSR: a lightweight fusion-enhanced adaptive perception network for image super-resolution

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Baiting Zhao,
  • Yingying Shang,
  • Xiaofen Jia,
  • Zhenhuan Liang,
  • Rui Hu

摘要

To address the challenges of high parameter counts and computational overhead in Swin Transformer-based image super-resolution methods, which hinder practical deployment, this paper proposes a lightweight dynamic adaptive fusion mixer network (DAFMixerSR). First, to enhance resource allocation flexibility, we devise a dynamic adaptive perception mixer module that combines a global predictor with streamlined contextual convolutions, effectively reducing computational complexity. Furthermore, an enhanced spatial-channel attentive (SCEA) mechanism is developed to fully integrate contextual information through spatial and channel attention, improving the efficiency of global-local feature extraction while significantly reducing parameters via depthwise separable convolutions. Additionally, a pixel-fusion convolutional feed-forward network (PFCFN) is introduced prior to the reconstruction module to further optimize restoration performance. Experimental results demonstrate that DAFMixerSR achieves an optimal balance between complexity and performance. For ×4 upscaling on the Set5 dataset, DAFMixerSR achieves PSNR improvements of 1.87 dB and 0.07 dB compared to classical SRCNN and state-of-the-art HSRNet, respectively, while requiring 46% fewer parameters than HSRNet.