Blind Image Denoising via Dependent Dirichlet Process Tree.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Traditional image denoising methods often assume known, homogeneous noise (e.g., Gaussian).
• Real-world noisy images frequently exhibit complex, unknown, and non-Gaussian noise distributions.
• Existing algorithms struggle with the inherent variability and unpredictability of noise in practical applications.

Purpose of the Study:

• To develop a novel blind image denoising algorithm capable of handling unknown and complex noise models.
• To propose a flexible noise modeling approach using a mixture of Gaussian distributions.
• To reformulate blind image denoising as a learning problem for improved performance.

Main Methods:

• Introduced a mixture of Gaussian distributions to model empirical noise flexibly.
• Formulated blind image denoising as a learning problem with a two-layer structural model for noisy patches.
• Proposed a novel Bayesian nonparametric prior, the Dependent Dirichlet Process Tree, to control model complexity.
• Derived a variational inference algorithm for parameter estimation and clean patch recovery.

Main Results:

• Demonstrated superior performance on both synthetic and real noisy images compared to existing methods.
• Successfully handled diverse and unknown noise models, showcasing algorithm robustness.
• Experimental results validate the algorithm's efficiency in practical image denoising tasks.

Conclusions:

• The proposed blind image denoising algorithm effectively addresses unknown and complex noise models.
• The use of mixture of Gaussians and Dependent Dirichlet Process Tree offers a flexible and powerful approach.
• The method shows significant potential for real-world applications requiring robust image restoration.