[Paper Review] A Derivative-Free Approach to Total Variation Regularization

Carsten Pontow, Otmar Scherzer|ArXiv.org|Nov 6, 2009

Numerical methods in inverse problems12 references3 citations

TL;DR

This paper introduces a derivative-free approach to total variation (TV) regularization by leveraging recent integral characterizations of the TV semi-norm via non-local double integrals. It establishes that minimizers of a sequence of approximating functionals converge to the minimizer of the original TV-regularized functional, enabling new numerical schemes and revealing a connection to bilateral filtering.

ABSTRACT

The goal of this paper is to present a novel approach for total variation regularization and Sobolev minimization, which are prominent tools for variational imaging. Thereby we use derivative free characterizations of the total variation semi-norm and Sobolev semi-norms of functions recently derived by Bourgain, Brézis, Mironescu and Dávila. Their analysis is to approximate the semi-norms of a function by singular integral operators. With this characterization we derive a series of novel regularization methods for total variation minimization which have as a novel feature a non-local double integral regularization term.

Motivation & Objective

• To develop a novel derivative-free framework for total variation regularization using recent integral characterizations of the TV semi-norm.
• To analyze the variational convergence of a sequence of approximating functionals based on singular integral operators.
• To derive new numerical schemes for TV minimization that avoid explicit derivative computation.
• To clarify the analytical foundation of existing numerical methods and reveal connections to bilateral filtering.
• To extend the approach to Sobolev spaces and provide a function-space setting for regularization without requiring differentiability of measures.

Proposed method

• Utilizes derivative-free characterizations of the TV semi-norm and Sobolev (1,p)-seminorms via double integrals involving difference quotients and approximation-to-identity kernels.
• Defines the regularization functional $ \mathcal{R}_n^p(f) = \int_\Omega \int_\Omega \frac{|f(x)-f(y)|^p}{|x-y|^p} \varphi_n(x-y) \,dx\,dy $, where $ \varphi_n $ are radially symmetric, non-negative, and converge to the Dirac delta at the origin.
• Constructs approximating functionals $ \mathcal{F}_n^p(f) = \frac{1}{2}\int_\Omega (f - f^\delta)^2 \,dx + \frac{\alpha}{K_{p,N}} \mathcal{R}_n^p(f) $, which converge to the original TV or Sobolev functional as $ n \to \infty $.
• Proves that each $ \mathcal{F}_n^p $ has a unique minimizer and that the sequence of minimizers converges to the minimizer of the limit functional $ \mathcal{F}^p $.
• Evaluates $ \mathcal{R}_n^1(f) $ explicitly for piecewise constant functions on a grid, yielding expressions in terms of absolute differences between neighboring pixel values.
• Applies the method to both radial and non-radial kernels (e.g., uniform kernel on a square), deriving closed-form expressions for the regularization term.

Experimental results

Research questions

• RQ1 Can the total variation semi-norm be approximated without using derivatives, using only integral operators?
• RQ2 Do the minimizers of the approximating functionals $ \mathcal{F}_n^p $ converge to the minimizer of the original TV or Sobolev functional?
• RQ3 What is the analytical structure of the non-local double integral regularization term for piecewise constant images?
• RQ4 How do different kernel functions (radial vs. non-radial) affect the resulting regularization functional?
• RQ5 Can this framework explain or unify existing numerical schemes like bilateral filtering?

Key findings

• The sequence of minimizers $ f_n $ of the approximating functionals $ \mathcal{F}_n^p $ converges to the unique minimizer of the original functional $ \mathcal{F}^p $, establishing variational convergence.
• For radial kernels, $ \mathcal{R}_n^1(f) $ evaluates to $ \frac{1}{3\pi n} \left( \sum_{i=2}^n \sum_{j=2}^n |a_{i,j} - a_{i-1,j-1}| + \cdots \right) $, with explicit coefficients for lateral and diagonal neighbors.
• For a uniform kernel on $ (-1/n, 1/n)^2 $, $ \mathcal{R}_n^1(f) $ yields $ \frac{1}{3} \frac{\sqrt{2}-1}{n} \sum |a_{i,j} - a_{i-1,j-1}| + \frac{1}{12} \frac{3\ln(\sqrt{2}+1) - 3\ln(\sqrt{2}-1) - 2(\sqrt{2}-1)}{n} \sum |a_{i,j} - a_{i,j-1}| $.
• The regularization term $ \mathcal{R}_n^1(f) $ is shown to be equivalent to a bilateral filtering operation in the limit, providing a theoretical link between TV denoising and non-local filtering.
• The method enables numerical schemes for TV minimization that do not require differentiability of the total variation measure, allowing a pure function-space formulation.
• The convergence of the approximating functionals is proven under standard assumptions on $ \Omega $ and $ f^\delta $, with $ \mathcal{F}_n^p $ having unique minimizers for all $ n $.