A Regularization Scheme Based on Gaussian Mixture Model for EM Data Inversion

Abstract

In this paper, we study parameter reconstruction from the perspective of probability, which is friendly to introduce prior information about the target region. The unknown contrast is assumed to follow Gaussian mixture model (GMM) and variational inference machinery is applied to realize the inversion. To decouple the contrast of different pixels, we consider the approximate posterior distribution from the perspective of optimization, and the inversion can be formulated as optimizing the combination of data misfit and prior information that works as the regularization.