Hierarchical Bayesian image restoration from partially known blurs.

Area of Science:

• Image processing and restoration
• Computational imaging
• Statistical signal processing

Background:

• Image degradation is a common problem in imaging systems.
• Partially known point-spread functions (PSFs) pose challenges for accurate image restoration.
• Parameter estimation for restoration filters has been a significant unresolved issue.

Purpose of the Study:

• To develop and present novel iterative algorithms for image restoration with partially known point-spread functions (PSFs).
• To address the problem of simultaneously estimating filter parameters and restoring degraded images.
• To provide solutions for parameter estimation in the context of regularized constrained total least-squares (RCTLS) and linear minimum mean square-error (LMMSE) filters.

Main Methods:

• Proposed two iterative algorithms using evidence analysis (EA) within a hierarchical Bayesian framework.
• Derived algorithms in the discrete Fourier transform (DFT) domain for computational efficiency.
• Presented an alternative approach to the expectation-maximization (EM) framework for LMMSE parameter estimation.

Main Results:

• The first algorithm's restoration step is nearly identical to the regularized constrained total least-squares (RCTLS) filter.
• The second algorithm's restoration step is identical to the linear minimum mean square-error (LMMSE) filter for this problem.
• Demonstrated computational efficiency for large images due to DFT domain implementation.

Conclusions:

• The proposed algorithms effectively address the simultaneous image restoration and parameter estimation problem for partially known PSFs.
• Provided a solution for parameter estimation in RCTLS filtering and an alternative to EM for LMMSE parameter estimation.
• The DFT-based approach ensures computational feasibility for practical applications in image restoration.