Regularization-parameter-free optimization approach for image deconvolution.

Area of Science:

• Computational Imaging
• Optimization Algorithms

Background:

• Image deconvolution is commonly an optimization problem with cost functions balancing data fidelity and image constraints.
• Selecting the regularization parameter for these cost functions typically requires empirical tuning, a time-consuming process for users.

Purpose of the Study:

• To present a novel image deconvolution framework, Mean Gradient Descent (MGD), that circumvents the need for regularization parameter tuning.
• To achieve a solution where data fidelity and image domain constraints are optimally balanced without manual parameter adjustments.

Main Methods:

• Developed a Mean Gradient Descent (MGD) methodology for image deconvolution.
• MGD progresses solutions by bisecting steepest descent directions of cost terms, balancing different objectives iteratively.
• Validated the approach using numerical simulations and experimental bright-field microscopy images.

Main Results:

• The MGD framework successfully performs image deconvolution without requiring regularization parameter selection.
• Demonstrated uniform deconvolution performance across datasets with varying noise levels.
• The method proved efficient and user-friendly in simulations and experimental data.

Conclusions:

• Mean Gradient Descent (MGD) provides an efficient, user-friendly alternative for image deconvolution, eliminating empirical parameter tuning.
• The MGD approach shows promise for various image deconvolution tools and potentially broader optimization problems.
• This method simplifies image restoration by automatically balancing competing cost functions.