Image restoration using space-variant Gaussian scale mixtures in overcomplete pyramids.

Area of Science:

• Computer Vision
• Signal Processing
• Image Restoration

Background:

• Bayes least squares-Gaussian scale mixtures (BLS-GSM) is a powerful image restoration method.
• Its effectiveness stems from local statistical descriptions of oriented pyramid coefficient neighborhoods.
• This method adapts to signal variance at each scale, orientation, and spatial location.

Purpose of the Study:

• To enhance the BLS-GSM model for improved image restoration.
• Introduce a coarser adaptation level for local signal covariance estimation.
• Apply the enhanced model to image denoising and deconvolution.

Main Methods:

• Formulated an enhanced BLS estimator using space-variant Gaussian scale mixtures (GSM).
• Incorporated a coarser adaptation level by using larger neighborhoods for covariance estimation within subbands.
• Applied the model to image deconvolution via global blur compensation followed by local adaptive denoising.

Main Results:

• The proposed method demonstrates significantly higher performance than the original BLS-GSM.
• Achieved superior results in both visual quality and L2-norm metrics for denoising and deconvolution.
• The enhanced method is model-based, noniterative, robust, and efficient.

Conclusions:

• The enhanced BLS-GSM model with space-variant GSM offers superior image restoration capabilities.
• The method provides a robust and efficient approach for both image denoising and deconvolution.
• Coarser adaptation levels improve local signal covariance estimation, leading to better restoration outcomes.