Seminar Details
2020-09-15 (14h) : Blind source separation of linear-quadratic near-separable mixtures
At Euler building (room A.002) + MS Teams
Organized by Mathematical Engineering
Speaker :
Christophe Kervazo (UMons)
Abstract :
In blind source separation (BSS), the data sets are assumed to stem from the mixing of elementary signals, called sources. Although the BSS methods assuming linear mixing models have known a wide range of successes, we will depart in this presentation from this setting and focus on the more advanced linear-quadratic (LQ) mixing models, which also include additional products of order two of the sources. Precisely, we here extend the near-separable nonnegative matrix factorization (NMF) framework to LQ mixings, enabling to design an efficient two-step unmixing algorithm coined SNPALQ + PP. The algorithm is shown to be robust to noise under discussed conditions. SNPALQ + PP furthermore exhibits good separation qualities in realistic hyperspectral unmixing numerical experiments.