Exact solutions for the total variation denoising problem of piecewise constant images in dimension one (Q2043449)

The present paper provides analytic results for the total variation denoising problem of piecewise constant functions in one dimension. The associated forward problem is given by \(f = Au + n\), where \(A\) is a bounded linear (blurring) operator, \(u\) is the uncorrupted (noise-free, sharp) image and \(n\) is additive noise. The inverse problem of denoising and/or deblurring is to recover \(u\) from \(f\). The present work is limited to the denoising problem (i.e., \(A\) in the definition of \(f\) is an identity operator). To recover \(u\), a variational approach is considered. The optimization problem consists of an \(L^p\)-distance functional with \(p > 1\) to measure the proximity of \(f\) and \(u\) and the total variation of \(u\). Instead of considering a fixed parameter \(\lambda\) (introduced as a dual variable for the constraint \(\|u-f\|_{L^p(\Omega)}^p\)), the analysis is carried out for a family of solutions \(u^\lambda\) starting with large values for \(\lambda\) and then obtaining solutions for decreasing \(\lambda\). Doing so allows for reducing the problem to one with a strictly convex functional, the minimizer of which is found by solving the associated Euler-Lagrange equations.