Blind deconvolution using generalized cross-validation approach to regularization parameter estimation.

Area of Science:

• Image processing
• Computational imaging
• Signal processing

Background:

• Blind deconvolution is crucial for image restoration, aiming to recover an image from its blurred and noisy version without prior knowledge of the blur kernel.
• Existing methods often struggle with accurate estimation of both the image and the blur simultaneously.
• Variational Bayesian methods with specific priors (Student's-t, total variation) have been explored but have limitations.

Purpose of the Study:

• To propose and present a novel algorithm for total variation (TV)-based blind deconvolution.
• To develop a method capable of estimating both the unknown image and the blur kernel concurrently.
• To enhance the performance and accuracy of blind deconvolution techniques.

Main Methods:

• An alternating minimization framework is employed to iteratively estimate the unknown image and blur.
• Generalized cross-validation (GCV) is utilized to update regularization parameters within the alternating minimization steps.
• The algorithm leverages total variation (TV) regularization for image prior and blur estimation.

Main Results:

• The proposed TV-based blind deconvolution algorithm successfully estimates both image and blur.
• Experimental results demonstrate superior performance compared to variational Bayesian blind deconvolution algorithms.
• The algorithm shows improved results over methods using Student's-t priors or a simple total variation prior.

Conclusions:

• The developed alternating minimization framework with GCV offers an effective solution for TV-based blind deconvolution.
• The proposed algorithm provides a robust and accurate approach for image and blur estimation in challenging scenarios.
• This method represents an advancement over current variational Bayesian techniques for blind deconvolution.