Semi-Blind Source Separation with Learned Physic-Driven Constraints

Rémi Carloni Gertosio (1, 2) , Jérôme Bobin (1, 2) , Fabio Acero (1)

1 CEA - CEA- Saclay
2 Université Paris-Saclay

Blind source separation (BSS) algorithms are unsupervised methods, which are the cornerstone of multispectral data analysis, allowing for a physically meaningful data decomposition. Being ill-posed inverse problems, BSS algorithms must rely on efficient regularization schemes to better distinguish between the sources and yield interpretable solutions. For that purpose, we investigate a semi-blind source separation approach in which we combine a projected alternate least-squares algorithm with a learning-based regularization scheme. In this article, we focus on constraining the mixing matrix to belong to a learned physics-based manifold; to that end, we propose to make use of the interpolatory autoencoder (IAE). Altogether, we show that this allows for an innovative BSS algorithm, with improved accuracy, which yet provides physically interpretable solutions. The proposed method, coined sGMCA, is tested on realistic multispectral astrophysical data in challenging scenarios involving strong noise, highly correlated spectra and unbalanced sources. The results highlight the significant benefit of the IAE prior to reduce the leakages between the sources, which allows a better disentanglement between the sources.

Domaines: Traitement du signal et de l'image [eess.SP]

Rémi Carloni Gertosio, Jérôme Bobin, Fabio Acero. Semi-Blind Source Separation with Learned Physic-Driven Constraints. 2021. ⟨hal-03270406⟩