Intitulé de l'offre : M/W Scientist for astronomical images data processing enhanced with machine learning (H/F)
Référence : UMR7326-ANAMEK-071
Nombre de Postes : 2
Lieu de travail : MARSEILLE 13
Date de publication : mardi 5 septembre 2023
Type de contrat : CDD Scientifique
Durée du contrat : 12 mois
Date d'embauche prévue : 1 novembre 2023
Quotité de travail : Temps complet
Rémunération : 2,932.85 to €4,669.28 gross/month, depending on experience.
Niveau d'études souhaité : Niveau 8 - (Doctorat)
Expérience souhaitée : Indifférent
Section(s) CN : Solar system and distant universe
The DOSSA project led by SpaceAble (https://spaceable.org/fr/) and conducted in partnership with Unistellar (https://www.unistellar.com/fr/) and the Laboratoire d'Astrophysique de Marseille has the ambition to bring to Europe a space monitoring capability with a global geographical coverage thanks to a new collaboration between a community of amateur astronomers, researchers and industrial networks. The challenge of the DOSSA (Decentralization of Space Situational Awareness) project is to develop French infrastructures that will increase our capacity to collect data and catalogue space objects while drastically reducing the cost of the data. This reduction is possible by relying on the already existing network of Unistellar telescopes, decentralized and with worldwide geographic coverage, and on participatory science by involving amateur astronomers in the project. This network is essential to intensify data collection and thus maximize the accuracy of orbital information, allowing to efficiently identify risks on exponentially growing catalogs of objects. The deployment of this network requires the development of image processing techniques enriched with Deep Learning models to increase telescope sensitivity and the ability to detect smaller satellites. This capability will provide the possibility to co-design smaller telescopes, thus less expensive and with a smaller footprint (1m2 versus 10m2), while guaranteeing the ability to detect objects of about 10 centimeters. The large number of collection points also ensures a more regular sampling of orbital parameters, and consequently a better ability to extrapolate trajectories and estimate conjunction risks. Deep Learning models are also essential to accelerate the calculations and ensure a rapid capacity to evaluate these risks on large catalogs.
WP1 - Creation of a simulator :
This task aims at creating a simulator of the Unistellar telescopes images including optical aberrations (design, atmosphere), mechanical aberrations (vibration, tracking error) and noise (sky background, detector noise, ..). This simulator will be verified thanks to the sky images delivered at the beginning of the project and will allow to feed the learning of the detection and astrometry models.
WP2 - Image denoising by auto-encoder :
This second work package deals with the problem of image denoising, in particular by applying methods based on auto-encoders. It will consist in (1) a state of the art of the currently available methods (see biblio below), (2) a classification of the methods according to the criteria of performance, execution speed and optimization/efficiency of the algorithms, (3) the implementation of the 3 most promising methods and (4) the benchmarking of these 3 methods using simulated images. Output: Auto-encoder based denoising code optimized for the study case.
WP3 - Image segmentation for optimal detection of satellite tracks :
This 3rd WP concerns the development of an algorithm based on "deep learning" approaches for the segmentation and optimal detection of satellite tracks in images. Based on the denoised images provided by WP1, the aim is now to identify and classify non-stellar objects in the images. The aim is to (1) review the state of the art of currently available methods based on cosmic ray detection methods (see bibliography below), (2) rank the methods according to performance, execution speed and algorithm optimization/efficiency criteria, (3) implement the 3 most promising methods and (4) benchmark these 3 methods using simulated images. Output: Segmentation and non-stellar object identification code in images, optimized for the case study.
Computing: Python, Matlab.
Optics, instrumentation, astronomy.
PhD in image processing, machine learning, instrumentation, optics, physics or astronomy.
Contexte de travail
The work will be carried at LAM, shared between the Research and Development Group and the Data center.
The duration of the contract is 12 months, renewable once.