Apprentissage automatique en grande dimension pour les applications d'imagerie radar

High dimension machine learning in radar imaging applications

• Auteur / Autrice : Cyprien Doz
• Direction : Jean-Philippe Ovarlez
• Type : Projet de thèse
• Discipline(s) : Sciences du traitement du signal et des images
• Date : Inscription en doctorat le 30/09/2019
• Etablissement(s) : université Paris-Saclay
• Ecole(s) doctorale(s) : École doctorale Sciences et technologies de l'information et de la communication (Orsay, Essonne ; 2015-....)
• Partenaire(s) de recherche : Laboratoire : SONDRA/CENTRALESUPELEC
• Partenaire(s) de recherche : Référent : CentraleSupélec (2015-....)

Mots clés

Mots clés libres

Images

Grandes dimensions

Classification

Matrices Aléatoires

Images

High dimension

Classification

Random Matrices

Résumé

1. Modélisation simple puis plus fine des images radar (avec différents modèles d'application), évaluation statistique des méthodes de clustering grâce aux avancées de la théorie des matrices aléatoires. 2. Analyse matricielle aléatoire appliquée à l'apprentissage statistique (clustering spectral, apprentissage semi-supervisé, transfer learning) afin de proposer des méthodes de clustering améliorées et d'en déduire leurs performances accessibles.

Clustering is an important technique in data mining which aims to group a collection of data items into clusters. This technique is also known as unsupervised classification of patterns (observations, data items or feature vectors) because it can be performed when no information is available concerning the membership of data items to predefined classes. The goal of this thesis is to develop new clustering methods that are accurate, robust and computationally efficient to automatically classify or recognize patterns, objects or classes in radar imaging application. The main novelty of the proposed approach is to leverage recent advances in large dimensional data clustering, notably through the help of random matrix theory. These methods particularly allow, for the first time, to obtain a theoretical grasp on the classification performances for various data models and thus to provide performance guarantees and limitations. These theoretical performances of classification algorithms are a major missing cornerstone of data-driven methods which are here available due to mathematical concentration arguments in large dimensional models.