Extending Nmf To Blindly Separate Linear-Quadratic Mixtures Of Uncorrelated Sources
Shahram Hosseini, Institut de Recherche en Astrophysique et Planétologie (IRAP), Toulouse University, UPS-OMP, CNRS, Toulouse
Yannick Deville, Institut de Recherche en Astrophysique et Planétologie (IRAP), Toulouse University, UPS-OMP, CNRS, Toulouse
Leonardo T. Duarte, The School of Applied Sciences, University of Campinas
Ahmed Selloum, Institut de Recherche en Astrophysique et Planétologie (IRAP), Toulouse University, UPS-OMP, CNRS, Toulouse

Abstract:
This paper proposes a new constrained method, based on non-negative matrix factorization, for blindly separating linear-quadratic (LQ) mixtures of mutually uncorrelated source signals when the sources and mixing parameters are all non-negative. The uncorrelatedness of the sources is used as a regularization term in the cost function. The main advantage of exploiting uncorrelatedness in this manner is that the inversion of the mixing model, which is a difficult task in the case of determined LQ mixtures, is not required, contrary to the classical LQ methods based on independent component analysis. Experimental results using artificial data and real-world chemical data confirm the effectiveness of our method.