A Preconditioned Version of a Nested Primal-Dual Algorithm for Image Deblurring

Stefano Aleotti,Marco Donatelli,Rolf Krause,Giuseppe Scarlato

2024-09-20

Abstract:Variational models for image deblurring problems typically consist of a smooth term and a potentially non-smooth convex term. A common approach to solving these problems is using proximal gradient methods. To accelerate the convergence of these first-order iterative algorithms, strategies such as variable metric methods have been introduced in the literature. In this paper, we prove that, for image deblurring problems, the variable metric strategy can be reinterpreted as a right preconditioning method. Consequently, we explore an inexact left-preconditioned version of the same proximal gradient method. We prove the convergence of the new iteration to the minimum of a variational model where the norm of the data fidelity term depends on the preconditioner. The numerical results show that left and right preconditioning are comparable in terms of the number of iterations required to reach a prescribed tolerance, but left preconditioning needs much less CPU time, as it involves fewer evaluations of the preconditioner matrix compared to right preconditioning. The quality of the computed solutions with left and right preconditioning are comparable. Finally, we propose some non-stationary sequences of preconditioners that allow for fast and stable convergence to the solution of the variational problem with the classical $\ell^2$--norm on the fidelity term.

Numerical Analysis

What problem does this paper attempt to address?

This paper attempts to solve the problems of optimization solution efficiency and precision in the image deblurring problem. Specifically, the article focuses on how to accelerate the convergence of proximal gradient methods through pre - processing strategies and reduce computational complexity.

### Problem Background

In the image deblurring problem, variational models are usually composed of a smoothing term and a convex term that may be non - smooth. A common solution method is to use proximal gradient methods. To accelerate the convergence of these first - order iterative algorithms, variable metric methods have been introduced in the literature. However, the traditional proximal gradient methods have two main drawbacks:
1. **Step - size selection problem**: If the selected step - size \(\alpha\) is too small, the convergence speed may slow down.
2. **Proximal operator calculation problem**: For some regularization terms (such as total variation), the proximal operator cannot be calculated explicitly.

### Paper Solutions

This paper proposes a Preconditioned Nested Primal - Dual Algorithm based on left pre - processing to solve the above problems. Specific improvements include:
1. **Re - interpretation of right pre - processing**: The author proves that for the image deblurring problem, the variable metric strategy proposed in the literature can be re - interpreted as a right pre - processing method.
2. **Introduction of left pre - processing**: An inexact left pre - processing version of the proximal gradient method is explored, and the convergence of the new iterative method is proved.
3. **Non - stationary pre - processing matrix sequence**: A non - stationary pre - processing matrix sequence is proposed, which allows fast and stable convergence to the solution of variational problems with the classical \(\ell_2\) - norm fidelity term.

### Main Contributions

- **Theoretical analysis**: It is proved that left pre - processing and right pre - processing are comparable in the number of iterations required to reach a predetermined tolerance, but left pre - processing requires less CPU time because it involves fewer pre - processing matrix evaluations.
- **Numerical results**: It is shown that left pre - processing and right pre - processing are comparable in the quality of the solution, and through the appropriate selection of non - smooth terms \(h\circ W\), pre - processing matrices \(P_n\) and extrapolation parameters \(\gamma_n\), this method can be aligned with standard algorithms such as ISTA, FISTA and ITTA.

### Mathematical Formulas

The main formulas involved in the paper include:

1. **Variational model**:
\[
\arg\min_{u\in\mathbb{R}^d}f(u)+h(Wu)
\]
where \(f:\mathbb{R}^d\rightarrow\mathbb{R}\) is a convex and smooth function, \(h:\mathbb{R}^{d'}\rightarrow\mathbb{R}\cup\{\infty\}\) is a convex and possibly non - smooth function, and \(W\in\mathbb{R}^{d'\times d}\) is a linear operator.

2. **Iterative formula of proximal gradient method**:
\[
u^{n + 1}=\text{prox}_{\alpha_n h\circ W}(u^n-\alpha_n\nabla f(u^n))
\]

3. **Iterative formula of left pre - processing**:
\[
\begin{cases}
\bar{u}^n=u^n+\gamma_n(u^n - u^{n - 1})\\
u^{n + 1}\approx\text{prox}_{P_n}^{\alpha_n h\circ W}(\bar{u}^n-\alpha_n P_n^{-1}\nabla f(\bar{u}^n))
\end{cases}
\]

Through these improvements, the paper provides a more efficient and stable optimization solution method for image deblurring.