Robust principal component analysis with intra-block correlation

Abstract

This paper proposes a novel optimization program for solving the Robust Principle Component Analysis (RPCA) problem, which decomposes a data matrix into a conventional low-rank part plus a particular block-sparse residual. This kind of block-sparse residual often scattered in the source signal scope as contaminants, and often existed in many practical applications, such as an ordinary imaging system, a Hyper Spectral Imaging system, EEG and MEG, and types of physiological signals. Different from most currently existing approaches, the study perceived especially a highly spatial correlation among the inner structure of the neighbouring pixels in this contiguously block-sparse residual. The high intra-block correlation is then introduced as prior information to deal the governing optimization problem. In order to enhance the block-sparsity and maintain the local smoothness simultaneously, a localized low-rank promoting method is introduced with a theoretical guarantee. An efficient solving algorithm is designed accordingly with a convergence analysis by adopting the classical Alternating Direction Method of Multipliers (ADMM) framework. In addition to the theoretical model derivation, several synthetic simulations together with a real application on image denoising experiment have been conducted to validate the proposed model. As expected, the models outperforms significantly the compared state-of-the-arts.