Convergence of an Iterative Method for Total Variation Denoising (Q4377535)

Consider the problem of reconstructing an unknown function \(u\), called the image, from data \(z\) satisfying \(z=u+\varepsilon \), where \(\varepsilon\) is error/noise in the data. Suppose \(u\in\Omega\subset\mathbb{R}^d\), \(d=1,2,\) or \(3\), which is bounded convex with Lipschitz continuous boundary \(\partial\Omega\). Suppose \(u\) is blocky, i.e. a nearly piecewise continuous function. A linear filtering techniques based on the total variation \(TV(u)=\int_{\Omega}| \nabla u| \) has been recognized to be very effective for denoising. The \(TV(u)\) is incorporated by considering the penalized least squares problem \[ \text{min}_u \textstyle{1\over 2}\| u-z\| _{L^2(\Omega)}+\alpha J(u), \] where \(J(u)\) is the replacement of \(TV(u)\) by a smooth approximation \(J(u)=\int_\Omega\sqrt{| \nabla | ^2+\beta^2}\) and \(\alpha\geq 0\) is a parameter controlling the trade-off between goodness of fit-to-the-data and variability in \(u\). The corresponding Euler-Lagrange operator equations are \[ (I+\alpha L(u))u=z\quad\text{with}\quad L(u)v=-\nabla \Biggl( {1\over\sqrt{| \nabla u| ^2+\beta^2}}\nabla v\Biggr). \] The fixed point iteration is then \[ u^{\nu +1}=(I+\alpha L(u^{\nu}))^{-1}z=F(u^{\nu}),\quad\nu=0,1,\dots\;. \] The paper analyzes the convergence of discrete approximations of fixed point iteration (FPI). The FPI is introduced in a finite-dimensional setting. Its global convergence (from any starting point) is established. The local convergence properties including rates and their dependence on various parameters (e.g. the \(\alpha\)-regularization and the \(\beta\)-smoothness parameters, and the discretization level) are examined.