Structure-Based Low-Rank Model With Graph Nuclear Norm Regularization for Noise Removal.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Nonlocal image representation methods like sparse coding and block-matching 3-D filtering are effective for low-level vision tasks.
• These methods extract nonlocal priors from similar image patches but can be inaccurate due to intensity-based grouping.

Purpose of the Study:

• To propose a structure-based low-rank model with graph nuclear norm regularization to overcome limitations of intensity-based patch grouping.
• To improve the accuracy and performance of image denoising algorithms.

Main Methods:

• Exploiting local manifold structure within image patches for patch grouping.
• Developing a graph nuclear norm regularization based on manifold structure information.
• Incorporating the regularization into a low-rank approximation model and solving it with a weighted singular-value thresholding algorithm.

Main Results:

• The proposed graph-based regularization is equivalent to a weighted nuclear norm.
• Experimental results demonstrate superior performance in additive white Gaussian noise removal and mixed noise removal compared to state-of-the-art methods.

Conclusions:

• The structure-based low-rank model with graph nuclear norm regularization offers a more accurate approach to image denoising.
• This method effectively addresses the inaccuracies associated with intensity-based patch grouping in nonlocal image representation.