Scholarlink

<p>Low-rank structures are commonly found in natural images, offering significant value in tasks such as image restoration and robust modeling. However, conventional low-rank models are mostly based on mathematical optimization models, which not only have high computational complexity but are also extremely sensitive to parameter settings. To more effectively exploit low-rank structures in images, we propose a Deep Unfolding Low-Rank Network, simplified DULRNet. By unfolding the optimization iterations of the Robust Principal Component Analysis (RPCA) model into a multistage neural network, DULRNet forms an interpretable and end-to-end trainable image restoration framework. During the update of key variables, DULRNet integrates learnable deep neural network modules with traditional optimization methods, enabling adaptive adjustment of important parameters. For example, the conventional Singular Value Thresholding (SVT) and soft-thresholding operators are replaced with convolutional neural network (CNN) modules, significantly enhancing the model’s adaptability. To further improve the recovery of global structures and edge details, the Total Generalized Variation (TGV) regularization is incorporated. Finally, extensive denoising experiments on both synthetic low-rank images and real-world datasets with various noise types demonstrate the superior restoration performance and strong structure-preserving ability. The source code for DULRNet will be released after publication at <a href="https://github.com/alcapa4120/DULRNet">https://github.com/alcapa4120/DULRNet</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

Deep unfolding low-rank network for image denoising

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Hengyou Wang,
  • Yang Sun,
  • Zhiming Liu,
  • Qiang He

摘要

Low-rank structures are commonly found in natural images, offering significant value in tasks such as image restoration and robust modeling. However, conventional low-rank models are mostly based on mathematical optimization models, which not only have high computational complexity but are also extremely sensitive to parameter settings. To more effectively exploit low-rank structures in images, we propose a Deep Unfolding Low-Rank Network, simplified DULRNet. By unfolding the optimization iterations of the Robust Principal Component Analysis (RPCA) model into a multistage neural network, DULRNet forms an interpretable and end-to-end trainable image restoration framework. During the update of key variables, DULRNet integrates learnable deep neural network modules with traditional optimization methods, enabling adaptive adjustment of important parameters. For example, the conventional Singular Value Thresholding (SVT) and soft-thresholding operators are replaced with convolutional neural network (CNN) modules, significantly enhancing the model’s adaptability. To further improve the recovery of global structures and edge details, the Total Generalized Variation (TGV) regularization is incorporated. Finally, extensive denoising experiments on both synthetic low-rank images and real-world datasets with various noise types demonstrate the superior restoration performance and strong structure-preserving ability. The source code for DULRNet will be released after publication at https://github.com/alcapa4120/DULRNet.