Compressive online decomposition of dynamic signals via n-l1 minimization with clustered priors
Host Publication: IEEE Statistical Signal Processing Workshop
Authors: H. Van Luong, N. Deligiannis, S. Forchhammer and A. Kaup
Publication Year: 2018
Number of Pages: 5
We introduce a compressive online decomposition via solving an nl1cluster-weighted minimization to decompose a sequence of datavectors into sparse and low-rank components. In contrast to conventionalbatch Robust Principal Component Analysis (RPCA)whichneeds to access full dataour method processes a data vector of thesequence per time instance from a small number of measurements.The n-l1 cluster-weighted minimization promotes (i) the structureof the sparse components and (ii) their correlation with multiplepreviously-recovered sparse vectors via clustering and re-weightingiteratively. We establish guarantees on the number of measurementsrequired for successful compressive decomposition under the assumptionof slowly-varying low-rank components. Experimental resultsshow that our guarantees are sharp and the proposed algorithmoutperforms the state of the art.