Variational Poisson Denoising via Augmented Lagrangian Methods

Christian Kanzow,Fabius Krämer,Patrick Mehlitz,Gerd Wachsmuth,Frank Werner

2024-06-21

Abstract:In this paper, we denoise a given noisy image by minimizing a smoothness promoting function over a set of local similarity measures which compare the mean of the given image and some candidate image on a large collection of subboxes. The associated convex optimization problem possesses a huge number of constraints which are induced by extended real-valued functions stemming from the Kullback--Leibler divergence. Alternatively, these nonlinear constraints can be reformulated as affine ones, which makes the model seemingly more tractable. For the numerical treatment of both formulations of the model (i.e., the original one as well as the one with affine constraints), we propose a rather general augmented Lagrangian method which is capable of handling the huge amount of constraints. A self-contained, derivative-free, global convergence theory is provided, allowing an extension to other problem classes. For the solution of the resulting subproblems in the setting of our suggested image denoising models, we make use of a suitable stochastic gradient method. Results of several numerical experiments are presented in order to compare both formulations and the associated augmented Lagrangian methods.

Optimization and Control

What problem does this paper attempt to address?

### Problems Addressed by the Paper

This paper aims to address the issue of image denoising, particularly in the context of Poisson noise. Specifically, the paper proposes a Variational Poisson Denoising (VPD) optimization problem and solves it using the Augmented Lagrangian Method.

#### Main Objectives
1. **Establish a Variational Poisson Denoising Model**: By defining a constrained optimization problem to estimate the true image \(\hat{u}\), where the constraints are based on local similarity measures to compare the given image and candidate images over multiple subregions.
2. **Propose a New Constraint Function**: Unlike traditional negative log-likelihood terms, the paper introduces a new constraint function \(HZ\), which is based on local likelihood ratio test statistics, ensuring that the candidate image is compatible with the observed data in each subregion.
3. **Develop Efficient Numerical Algorithms**: For this non-smooth convex optimization problem with numerous constraints, an Augmented Lagrangian Method is proposed, and its convergence is proven. Additionally, a method to transform nonlinear constraints into affine constraints is considered, making the problem easier to solve.

#### Methodological Contributions
1. **Variational Poisson Denoising Optimization Problem (VPD)**: The paper defines an optimization problem by minimizing a smoothness-promoting function \(f(u)\) while satisfying a set of constraints defined by local likelihood ratio test statistics of the Poisson distribution.
2. **Augmented Lagrangian Method**: To solve this problem with numerous constraints, the paper proposes an Augmented Lagrangian Method and provides a self-contained, derivative-free global convergence theory.
3. **Numerical Experiment Results**: Multiple numerical experiments validate the effectiveness and performance of the two different forms of the model and their corresponding Augmented Lagrangian Methods.

In summary, the main purpose of this paper is to propose a new Variational Poisson Denoising model and solve it using an efficient Augmented Lagrangian Method, thereby achieving better denoising results when dealing with Poisson noise in images.