Abstract
We propose a PDE-constrained optimization approach for the determination of noise distribution in total variation (TV) image denoising. An optimization problem for the determination of the weights correspondent to different types of noise distributions is stated and existence of an optimal solution is proved. A tailored regularization approach for the approximation of the optimal parameter values is proposed thereafter and its consistency studied. Additionally, the differentiability of the solution operator is proved and an optimality system characterizing the optimal solutions of each regularized problem is derived. The optimal parameter values are numerically computed by using a quasi-Newton method, together with semismooth Newton type algorithms for the solution of the TV-subproblems.
Introduction
In the field of computer vision, image denoising remains a fundamental challenge. The objective is to recover a clean image $u$ from a noisy observation $f$. Total Variation (TV) denoising has long been a standard approach due to its ability to preserve edges while smoothing noise. However, the effectiveness of TV denoising is highly dependent on the choice of the "fidelity term," which must match the underlying noise distribution (e.g., Gaussian, Poisson, or impulse noise).
In practice, the specific noise characteristics of a given imaging system are often unknown or mixed, leading researchers to rely on manual tuning or heuristic guesses. This paper, authored by Juan Carlos de los Reyes and Carola-Bibiane Schönlieb, introduces a sophisticated mathematical framework to automate this process. By treating noise model selection as a bilevel optimization problem, the authors enable the system to "learn" the optimal noise distribution and regularization parameters directly from data.
The Problem: Automatic Model Selection
Standard TV denoising involves minimizing a functional of the form: $$J(u) = \text{Regularization}(u) + \lambda \cdot \text{Fidelity}(u, f)$$
The mathematical form of the fidelity term changes based on the noise type:
- Gaussian Noise: Requires an $L^2$ norm.
- Poisson Noise: Requires a Kullback-Leibler distance.
- Impulse (Salt-and-Pepper) Noise: Requires an $L^1$ norm.
Selecting the wrong model or an incorrect weight $\lambda$ results in either oversmoothing or residual noise. The authors address this by proposing a model that considers a weighted combination of multiple fidelity terms, where the weights themselves are variables to be optimized.
Methodology: Bilevel Optimization
The core of the proposed approach is a Bilevel Optimization strategy. This can be conceptualized as a hierarchical system where one optimization problem is nested within another:
- The Lower-Level Problem (Denoising): This is the variational task of finding the best image $u$ given a set of noise weights. The fidelity term is a weighted sum representing different potential noise models.
- The Upper-Level Problem (Learning): This level seeks to find the optimal weights (parameters) that minimize the discrepancy between the denoised output and a known "ground truth" image from a training set.
Computational Techniques
To solve this complex structure, the authors employ several advanced mathematical tools:
- Huber Regularization: Because the standard TV functional is non-differentiable, the authors use Huber smoothing. This makes the functions differentiable, allowing for the use of gradient-based optimization methods.
- PDE-Constrained Optimization: The denoising process is treated as a constraint—specifically a Variational Inequality—within the broader parameter-learning problem.
- Algorithmic Solvers: The authors utilize a Quasi-Newton (BFGS) method for the upper-level parameter search and a Semismooth Newton algorithm to efficiently solve the lower-level TV-subproblems.
Experimental Validation
The framework was tested across several scenarios to evaluate its robustness and accuracy:
- Synthetic Proof-of-Concept: Using "toy" images, the authors demonstrated that the optimizer could successfully identify Gaussian and impulse noise parameters.
- Medical Imaging (MRI): The method was applied to brain scan images with varying noise levels (3% and 9%). This demonstrated a potential real-world application: "calibrating" the denoising parameters for specific medical hardware by using a high-quality phantom scan as ground truth.
- Mixed Noise Scenarios: In tests involving a combination of Gaussian and Poisson noise (using an image of a plane wing), the algorithm converged rapidly—often within 12 iterations—and proved robust even when the test images were slightly shifted or displaced.
Key Insights
One of the most significant findings is the automatic sparsity of the solution. The optimization process naturally drives the weights of "incorrect" noise models toward zero. For instance, if an image contains only Gaussian noise, the weights for Poisson and Impulse terms will diminish, effectively performing automatic model selection.
Furthermore, the paper provides rigorous mathematical grounding. Unlike many empirical "black box" machine learning models, this work proves the existence of an optimal solution and ensures the consistency of the regularization approach, providing a reliable theoretical foundation for the numerical results.
Limitations and Future Directions
While powerful, the approach has limitations. The "learning" phase requires a training set with ground truth images, which are not always available in fields like deep-space photography or certain real-time sensing applications. Additionally, bilevel optimization is computationally intensive, as every step in the parameter search requires solving a complete denoising problem.
Future work may focus on reducing this computational overhead and addressing the theoretical challenges of proving continuity in the "control-to-state" map for pure, non-regularized TV functionals.
Summary updated