Applied and Interdisciplinary Mathematics Seminar

University of Michigan

Fall 2007
Tuesday, 2 October, 3:10-4:00pm, 4088 East Hall

On the regularity and curvature properties of level sets of minimizers for denoising models using total variation regularization

William K. Allard

Duke University


Abstract

Suppose s is a noisy grayscale image which we wish to denoise; mathematically, s will be a bounded real valued function on R^2. One way to do this is as follows. Let \gamma: R --> [0,\infty) be zero at zero, positive away from zero and convex. Let \epsilon >0 be a ``smoothing'' parameter which we will tune appropriately and let F_\epsilon(f) = \epsilon TV(f) + \int \gamma(f(x)-s(x)) dx for f: R^2 --> [0,\infty); here TV(f) is the total variation of f. Minimizers of F_\epsilon are used to denoise s. In this talk we will describe some notable geometric properties of minimizers and will provide some interesting examples. It turns out that these results rest on the theory of area minimization which has been in the literature for over 35 years.