Authors: Z. Chen (Northwestern University, Illinois, USA), R. Molina (Universidad de Granada, Spain), A. Katsaggelos (Northwestern University, USA)
[Projet Page][Publications]
Chapter Description
In this chapter, we consider a generalization of the original RPCA problem, where a linear transformation through the use of a known measurement matrix, is applied to the outlier-corrupted data. The goal is to estimate the outlier amplitudes given the transformed observation. This problem stems from several practical scenarios, which we will discuss in detail shortly. A regularization based algorithm, which requires the manual tuning of its parameters, was proposed to solve this problem. In this work, we propose a variational Bayesian based approach that provides approximate posterior distributions of all the model unknowns. Experiments using real life datasets as well as computer simulations show performance improvement of the proposed algorithm over its regularization based counterpart.