Offre en lien avec l’Action/le Réseau : – — –/– — –
Laboratoire/Entreprise : Laboratoire des Signaux et Systèmes
Durée : 3 ans
Contact : francois.orieux@l2s.centralesupelec.fr
Date limite de publication : 2021-04-22
Contexte :
More details can be found here https://pro.orieux.fr/files/data-fusion.pdf
The objective of the subject is to develop new methods, notably inspired by machine learning approach, for fusion of heterogenous data, hyper and multispectral images in particular. It takes place in the international James Webb Space Telescope (JWST) project, the most ambitious space telescope scheduled for launch in October 2021 and which will carry four instruments with unprecedented hyperspectral and multispectral observing capabilities.
Hyperspectral and multispectral imaging are ubiquitous in many observation modalities like Earth observations (like the European Copernicus project), astronomy, medical imaging, material analysis. Thus the work carried out during the PhD should also be applicable to many other modalities.
This thesis is part of the scientific project of the joint laboratory LAB4S2 (Innovative Laboratory for Space Spectroscopy) between the IAS and the company ACRI-ST (https://www.acri- st.fr/fr). The objective of LAB4S2 is to develop fusion methods for the joint exploitation of hyperspectral and multispectral data obtained, on the one hand, by the mid-infrared instrument of the JWST (MIRI), and, on the other hand, by the instruments on board the Sentinel missions of the Copernicus project.
Sujet :
The aim of the thesis is to develop efficient algorithms for joint processing:
– of multispectral data obtained with imager over a large field of view and broadband filters ;
– with hyperspectral data with high spectral resolution but for small fields of view.
This problem, which can be likened to a data fusion problem, has been addressed in Earth observations (with so-called pansharpening methods), in a context where the effects due to the observing instruments (limited spatial resolution and insufficient spatial sampling) are negligible.
The problem can be solved by minimization of a loss function with explicit data models for multispectral (imaging) and hyperspectral (spectroscopic) instruments. This approach solves the problem of measurement heterogeneity by relying on a “virtual” instrument that combines imaging and spectroscopy.
The above-mentioned method is model-driven. However, the model can be inefficient is some case, when the degradations due to the instruments, like spatial blurring, are important for instance. A common approach is to add a regularizer. This approach can have some limitations. In particular, the prior models are often ad hoc and the compromise between the different sources of information must be tuned. During this work we will explore alternative methods based on machine learning to construct more adapted prior. The challenge here will be the lake of a big database necessary for supervised approach. Therefore we will explore new methods based on small databases like Few-Shot approach. This approach consists of adapting trained DNN with existing databases (like ImageNet) with another database that have only a few examples.
Profil du candidat :
A professional attitude is expected.
Formation et compétences requises :
The candidate must have a M2 degree in signal processing, data science, or machine learning. Knowledge in physics or optics is appreciated.
Adresse d’emploi :
Laboratoire des Signaux et Systèmes (Université Paris-Saclay)
3 rue Joliot-Curie, 91192 Gif-sur-Yvette
Document attaché : 202102221056_data-fusion.pdf