Blind nonlinear source separation using DCT modeling
Marc Ibanyez, Ana I. Pérez-Neira, Miguel Angel Lagunas · 2024
The challenge lies in the blind separation of sources from a non-linear mixture, specifically focusing on non-linear mutual coupling mixtures. In this study, the authors revisit the pioneering method introduced by C. Jutten and G. Herault (J-H) for source separation in linear mixtures and propose an architecture incorporating adaptive non-linearities. The mixing architecture assumes that each original source at the observations is affected by the addition of a non-linear function of the other independent source, exemplifying non-linear mutual coupling. Therefore, the paper is restricted to a non-linear structure assumed for non-linear mixtures. The objective is to demonstrate the expressivity of the trigonometric approach, based on the Discrete Cosine Transform (DCT), in modeling non-linear systems. The resulting separation network and algorithm mimic the original J-H method by identifying the non-linear blocks of the mix and separating the two independent sources with extremely low complexity.