Image Denoising Based on Nonlocal Bayesian Singular Value Thresholding and Stein's Unbiased Risk Estimator

Area of Science:

• Image processing
• Computer vision
• Signal processing

Background:

• Nonlocal image denoising methods like singular value thresholding (SVT) and nuclear norm minimization (NNM) are sensitive to noise variance estimation.
• Existing methods often assume known noise variance or require separate estimation steps, leading to error propagation and performance degradation.
• Least squares estimation, common in these methods, can result in high mean-squared error (MSE) and limitations with missing data or outliers.

Purpose of the Study:

• To address the limitations of existing SVT/NNM-based denoising methods.
• To propose a hybrid denoising model that accurately estimates noise variance and reduces artifacts.
• To improve the performance of image denoising, especially in the presence of noise and potential data imperfections.

Main Methods:

• A hybrid denoising model combining variational Bayesian inference and Stein's unbiased risk estimator (SURE).
• The first step involves variational Bayesian SVT for low-rank approximation, simultaneously denoising and estimating noise variance.
• The second step modifies conventional SURE SVT and its divergence formulas for rank-reduced eigen-triplets to eliminate residual artifacts.

Main Results:

• The proposed hybrid BSSVT method demonstrates superior performance in recovering true images compared to state-of-the-art techniques.
• Simultaneous noise removal and noise variance estimation achieved through variational Bayesian SVT.
• Effective removal of residual artifacts using modified SURE SVT for rank-reduced eigen-triplets.

Conclusions:

• The hybrid BSSVT method offers a robust solution for image denoising by overcoming the limitations of traditional SVT/NNM approaches.
• Accurate noise variance estimation and artifact reduction are key contributions of the proposed model.
• The method shows significant improvements in image recovery, outperforming existing state-of-the-art denoising techniques.