Scholarlink

<p>In image denoising, neural network models based on the UNet architecture have become the mainstream due to their excellent performance and strong reconstruction capabilities. However, their heavy reliance on paired data and high computational costs limit their deployment and application on embedded devices. This paper adopts the Blind2Unblind (B2U) self-supervised training framework and proposes a lightweight U-Net-based denoising network with improved computational efficiency. This network is based on the UNet architecture, enhancing multi-scale denoising capabilities through multiple convolutional blocks and branch feature fusion, and introducing global attention by combining depthwise separable convolution, short skip connections, and GCBlock. To achieve lightweight, the softmax in GCBlock is replaced with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {L}^{-\text {1}}\text {ReLU}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mtext>L</mtext> <mrow> <mo>-</mo> <mtext>1</mtext> </mrow> </msup> <mtext>ReLU</mtext> </mrow> </math></EquationSource> </InlineEquation> to reduce computational costs. Experiments are conducted on infrared synthetic noise, SIDD, and synthetic datasets with Poisson/Gaussian noise. The results show that the proposed network significantly reduces model complexity while maintaining competitive denoising performance. Additional comparisons with recent representative denoising models further demonstrate that the proposed method offers a favorable trade-off between denoising quality and computational complexity, making it a denoising model with high computational efficiency and the capability of deploying on edge devices.</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

MSAU-Net: A Lightweight Self-Supervised Image Denoising Network with ReLU-Based Global Context Attention

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Jiye Huang,
  • Jiahan Yang,
  • Lei Tang,
  • Xin Liu,
  • Zhiqiang Cao,
  • Xiwei Huang

摘要

In image denoising, neural network models based on the UNet architecture have become the mainstream due to their excellent performance and strong reconstruction capabilities. However, their heavy reliance on paired data and high computational costs limit their deployment and application on embedded devices. This paper adopts the Blind2Unblind (B2U) self-supervised training framework and proposes a lightweight U-Net-based denoising network with improved computational efficiency. This network is based on the UNet architecture, enhancing multi-scale denoising capabilities through multiple convolutional blocks and branch feature fusion, and introducing global attention by combining depthwise separable convolution, short skip connections, and GCBlock. To achieve lightweight, the softmax in GCBlock is replaced with \(\text {L}^{-\text {1}}\text {ReLU}\) L - 1 ReLU to reduce computational costs. Experiments are conducted on infrared synthetic noise, SIDD, and synthetic datasets with Poisson/Gaussian noise. The results show that the proposed network significantly reduces model complexity while maintaining competitive denoising performance. Additional comparisons with recent representative denoising models further demonstrate that the proposed method offers a favorable trade-off between denoising quality and computational complexity, making it a denoising model with high computational efficiency and the capability of deploying on edge devices.