Image modeling and denoising with orientation-adapted Gaussian scale mixtures

Area of Science:

• Image Processing
• Statistical Modeling
• Computer Vision

Background:

• Existing image denoising methods often struggle with spatially varying signal characteristics.
• Multiscale image representations require sophisticated models to capture local coefficient behaviors.
• Gaussian scale mixtures have been used but can be limited in adapting to local signal properties.

Purpose of the Study:

• To develop a novel statistical model for describing spatially varying local neighborhoods of coefficients in multiscale image representations.
• To create an optimal Bayesian least squares estimator for image denoising based on the proposed model.
• To evaluate the denoising performance of the new method against established techniques.

Main Methods:

• Modeled local neighborhoods as multivariate Gaussian density samples, modulated and rotated by hidden random variables.
• Incorporated a third hidden variable to switch between oriented and non-oriented Gaussian processes for adaptability.
• Developed an optimal Bayesian least squares estimator for image denoising.

Main Results:

• The proposed statistical model effectively adapts to local signal amplitude, orientation, and orientedness.
• Simulations demonstrated significant performance improvements in image denoising compared to previous methods.
• The Bayesian least squares estimator achieved superior results over Gaussian scale mixture approaches.

Conclusions:

• The developed statistical model provides a robust framework for representing complex local image structures.
• The proposed Bayesian denoising method offers a significant advancement in image quality restoration.
• This approach demonstrates enhanced adaptability and effectiveness for image denoising tasks.