Publication Details
Overview
 
 
Bart Truyen, Cristina Boca, Ronny Hoffmann
 

Chapter in Book/ Report/ Conference proceeding

Abstract 

We disclose a novel inversion algorithm for electrical impedance tomography in two dimensions, based on a consistent subspace interpretation of the underlying severely ill-posed inverse problem. Whereas conventional output-least-squares algorithms can be regarded as minimizing a specific error norm, solutions are recovered here as the minimizers of a closely related, equality-constrained residual norm problem. The resulting sparse problem formulation defines sets of discrete subspaces, to which admissible solutions necessarily are confined. Unlike for output-least-squares, these subspaces no longer are continuous, but instead will assume a discontinuous form, whereby noise perturbations systematically are restrict in their effects. We formulate an efficient Gauss-Newton iteration for finding a solution to the nonlinear problem, and show that this gives rise to a sequence of sparse matrix problems, for which a substantial better conditioning can be demonstrated than typically observed for output-least-squares. A sparse QR factorization is developed that takes full advantage of the block angular matrix structures.

Reference