New approaches for non-linear blind source separation, with application to remote sensing data.
PhD offer at Télécom Paris.
The whole subject can be accessed at https://sites.google.com/view/christophekervazo/. For additional informations, please contact email@example.com.
Blind source separation, Signal and Image Processing, Remote sensing and hyperspectral imaging, Nonnegative Matrix Factorization, Provable Machine learning
Blind source separation (BSS) is a powerful machine learning paradigm with a wide range of applications such as remote sensing and biomedical imaging. Generally speaking, BSS aims at decomposing a given data set into unknown elementary signals to be recovered, generally referred to as the sources.
Because it is simple and easily interpretable, much attention has been dedicated to the linear mixing model (LMM), in which the mixing process is assumed to be linear. Building on the LMM and some additional assumptions, previous works have proposed several provably robust BSS algorithms, that is algorithms which are proved to recover the source signals at the origin of the considered data set even in the presence of noise.
On the other hand, in various BSS applications such as hyperspectral unmixing, the LMM is unfortunately only a first-order approximation of non-linear mixing processes. The goal of this PhD is thus to investigate several promising research paths to develop provably robust non-linear BSS methods, as very few algorithms have been proposed in this direction.
The developed methods will be applied to remote sensing hyperspectral unmixing. This field has a strong interest, in particular for urban development or vegetation monitoring in the context of global warming.
Please see more details at: https://sites.google.com/view/christophekervazo/
The candidate should have a Master 2 degree (or equivalent) and an excellent academic curriculum. He/she should have a good knowledge in signal/image processing and mathematics (especially, linear algebra). Knowledge in convex optimization is a plus. Ideally, Matlab programming language should be mastered.
The candidate will acquire an expertise in signal processing (in particular, of multi-valued data), which is valuable in many fields : remote sensing, astrophysics, text-mining...
The project will be conducted under the supervision of Christophe Kervazo and Florence Tupin, within the IMAGES group, at the Telecom Paris engineering school. The work is expected to be performed in collaboration with Nicolas Gillis, who is a reknown expert in the field of provably robust BSS.
Candidates must take contact and send a detailed CV to Christophe Kervazo (firstname.lastname@example.org)
(c) GdR 720 ISIS - CNRS - 2011-2020.