The equivalence of half-quadratic minimization and the gradient linearization iteration

Area of Science:

• Image processing
• Computational mathematics
• Computer vision

Background:

• Image restoration often involves minimizing complex cost functions with edge-preserving regularization terms.
• Nonconvex regularization terms, while effective for edge preservation, lead to slow and computationally intensive solutions.
• Half-Quadratic (HQ) minimization was introduced to address these computational challenges in image reconstruction.

Purpose of the Study:

• To demonstrate the equivalence between multiplicative Half-Quadratic (HQ) minimization and a basic iterative gradient descent method.
• To simplify and accelerate image restoration processes that utilize nonconvex regularization.
• To explore the connections between HQ minimization and other optimization techniques.

Main Methods:

• Reformulation of the optimization problem using an augmented cost function (HQ minimization).
• Analysis of the iterative steps of multiplicative HQ minimization.
• Comparison of HQ minimization iterations with those of a linearized gradient descent approach.
• Investigation of relationships with quasi-Newton and generalized Weiszfeld methods.

Main Results:

• Multiplicative HQ minimization is mathematically equivalent to a simple iterative gradient descent method.
• Both methods produce identical iteration sequences for image restoration.
• Established straightforward connections between HQ minimization and other established optimization algorithms.

Conclusions:

• The computational complexity of image restoration using HQ minimization can be reduced by recognizing its equivalence to simpler iterative methods.
• This equivalence offers a more accessible understanding and implementation of advanced image restoration techniques.
• The findings facilitate further research into efficient and effective image reconstruction algorithms.