Robust multichannel blind deconvolution via fast alternating minimization.

Area of Science:

• Image processing
• Computational imaging
• Optimization theory

Background:

• Blind deconvolution is an ill-posed problem for estimating images and blur simultaneously.
• Multichannel (MC) blind deconvolution improves problem conditioning using multiple images.
• Existing MC methods can be sensitive to noise and inaccuracies in blur estimation.

Purpose of the Study:

• To develop a more robust and stable multichannel blind deconvolution algorithm.
• To address limitations of existing methods concerning noise and large/overestimated blurs.
• To provide a computationally efficient solution for blind image deconvolution.

Main Methods:

• Formulating blind deconvolution as an l(1)-regularized optimization problem.
• Alternately optimizing with respect to the image and blur parameters.
• Employing variable splitting and augmented Lagrangian methods for constrained optimization.
• Implementing the solution efficiently in the Fourier domain.

Main Results:

• The proposed method demonstrates enhanced robustness to noise.
• Improved stability is achieved even with large blurs or overestimated blur sizes.
• Rapid convergence of the optimization algorithm was observed.
• Successful deconvolution of real digital camera images was demonstrated.

Conclusions:

• The enhanced multichannel blind deconvolution method offers improved performance and stability.
• The algorithm is effective for both synthetic and real-world image deconvolution tasks.
• The Fourier domain implementation ensures computational efficiency.