Les stages

PDF

Machine learning pour l'imagerie en astrophysique
Machine Learning and Image Deblending for Astrophysical Data

Spécialité

Traitement d'image

Niveau d'étude

Bac+5

Formation

Master 2

Unité d'accueil

Candidature avant le

20-05-2018

Durée

4 mois

Poursuite possible en thèse

oui

Contact

STARCK Jean-luc
+33 1 69 08 57 64

Résumé/Summary

Blending of astrophysical sources has a significant impact on the measurement of the morphological properties of galaxies. It is therefore essential to develop effective and reliable methods for identifying blended sources in survey data. The goal of this internship is to develop a machine learning technique for this task
Blending of astrophysical sources has a significant impact on the measurement of the morphological properties of galaxies. It is therefore essential to develop effective and reliable methods for identifying blended sources in survey data. The goal of this internship is to develop a machine learning technique for this task

Sujet détaillé/Full description

Context
Upcoming astrophysical surveys such as CFIS 1 and Euclid 2 aim to constrain cosmological parameters
using properties derived from galaxy images, in particular their shapes via weak gravitational lensing.
However, blending of sources (i.e. the overlap of extended objects) has a significant impact on the
measurement of the morphological and structural properties of galaxies. It is therefore essential to
develop effective and reliable methods for identifying blended sources in survey data and establishing
appropriate means of dealing with them.
Machine learning techniques have been show to be incredibly successful when applied to complex
classification problems (see e.g. Kotsiantis 2007), while signal processing techniques, such as sparsity,
have been shown to be extremely effective at deblending multi-band images (Joseph et al. 2016). The
combination of these tools may help in reducing the bias on galaxy properties introduced by blended
sources.
Outline of project objectives
The internship will be broadly divided into the following main blocks and objectives:
1. Get familiarised with the work of a previous internship, which compared the performance of SExtractor
(Bertin & Arnouts 1996) with machine learning methods for identifying blended sources in
monochromatic images.
2. Extend the existing applications of machine learning tools for identifying blended sources to simulations
of multi-band data and to real observed images.
3. Develop new techniques for dealing with blended sources using machine learning and/or signal
processing tools.
4. Interact with other members in CosmoStat to gauge the impact of the newly developed deblending
scheme on projects such as CFIS.
Candidate
The candidate should be a Master 2 (or equivalent) student with background in either physics/astrophysics
or applied maths/signal processing/data science. Knowledge machine learning methods would be a plus.
Experience coding in Python is not required, but would be advantageous.
Internship
The internship will take place in the CosmoStat laboratory, under the supervision of Jean-Luc Starck
and Samuel Farrens.
• Deadline for applications: February 28th, 2018.
• Contact: Samuel Farrens (samuel.farrens@cea.fr).
• Duration: 4-6 months.
• Possibility to continue on for a PhD: Yes.
1http://www.cfht.hawaii.edu/Science/CFIS/
2https://www.euclid-ec.org/
1
References
Bertin, E. & Arnouts, S. 1996, Astronomy and Astrophysics Supplement, 117, 393
Joseph, R., Courbin, F., & Starck, J.-L. 2016, Astronomy & Astrophysics, 589, A2
Kotsiantis, S. 2007, 31, 249

Context
Upcoming astrophysical surveys such as CFIS 1 and Euclid 2 aim to constrain cosmological parameters
using properties derived from galaxy images, in particular their shapes via weak gravitational lensing.
However, blending of sources (i.e. the overlap of extended objects) has a significant impact on the
measurement of the morphological and structural properties of galaxies. It is therefore essential to
develop effective and reliable methods for identifying blended sources in survey data and establishing
appropriate means of dealing with them.
Machine learning techniques have been show to be incredibly successful when applied to complex
classification problems (see e.g. Kotsiantis 2007), while signal processing techniques, such as sparsity,
have been shown to be extremely effective at deblending multi-band images (Joseph et al. 2016). The
combination of these tools may help in reducing the bias on galaxy properties introduced by blended
sources.
Outline of project objectives
The internship will be broadly divided into the following main blocks and objectives:
1. Get familiarised with the work of a previous internship, which compared the performance of SExtractor
(Bertin & Arnouts 1996) with machine learning methods for identifying blended sources in
monochromatic images.
2. Extend the existing applications of machine learning tools for identifying blended sources to simulations
of multi-band data and to real observed images.
3. Develop new techniques for dealing with blended sources using machine learning and/or signal
processing tools.
4. Interact with other members in CosmoStat to gauge the impact of the newly developed deblending
scheme on projects such as CFIS.
Candidate
The candidate should be a Master 2 (or equivalent) student with background in either physics/astrophysics
or applied maths/signal processing/data science. Knowledge machine learning methods would be a plus.
Experience coding in Python is not required, but would be advantageous.
Internship
The internship will take place in the CosmoStat laboratory, under the supervision of Jean-Luc Starck
and Samuel Farrens.
• Deadline for applications: February 28th, 2018.
• Contact: Samuel Farrens (samuel.farrens@cea.fr).
• Duration: 4-6 months.
• Possibility to continue on for a PhD: Yes.
1http://www.cfht.hawaii.edu/Science/CFIS/
2https://www.euclid-ec.org/
1
References
Bertin, E. & Arnouts, S. 1996, Astronomy and Astrophysics Supplement, 117, 393
Joseph, R., Courbin, F., & Starck, J.-L. 2016, Astronomy & Astrophysics, 589, A2
Kotsiantis, S. 2007, 31, 249

Mots clés/Keywords

machine learning
astrophysique

Compétences/Skills

machine learning teachniques

Logiciels

Python
Retour en haut