Applied and Interdisciplinary Mathematics Seminar

University of Michigan

Fall 2007
Monday, 22 October, 4:10-5:00pm, 3088 East Hall

The effectiveness of Bregman iteration as applied to compressed sensing and image restoration

Stanley J. Osher

UCLA


Abstract

Bregman iterative regularization (1967) was introduced by Osher, Burger, Goldfarb, Xu and Yin as a device for improving TV based image restoration (2004) and was used by Xu and Osher in (2006) to analyze and improve wavelet shrinkage. In recent work by Yin, Osher, Goldfarb and Darbon we devised simple and extremely efficient methods for solving the basis pursuit problem which is used in compressed sensing. A linearized version done by Osher, Dong and Yin requires two lines of MATLAB code and is remarkably efficient. This means we rapidly and easily solve the problem: minimize over u in R^n {myu||u||_1 +1/2 (||Au-f||_2)2} for a given k by n matrix A, with k much less than n and f in R^k By some beautiful results of Candes, Tao, Donoho and collaborators, this L1 minimization gives the sparsest solution u, under reasonable assumptions.