Distance metrics comparison for multispectral image processing

Sujets de thèse 2013

Intitulé de la thèse
Distance metrics comparison for multispectral image processing

Publication du sujet sur le site de l’ABG : OUI
Nature du financement : Financement institutionnel, Contrat Doctoral, Financement régional, Contrats université sur projets,)
Domaine de compétences principal (pour l’ABG) : Informatique, électronique
Domaine de compétences secondaire (pour l’ABG) : Sciences pour l’Ingénieur
Spécialité de doctorat : Image Signal et Automatique

Lieu de travail
Laboratoire XLIM-SIC, équipe ICONES, SP2MI, Futuroscope

Cotutelle avec The Norwegian Colour and Visual Computing Laboratory,
Gjøvik University College, Norvège

Date Limite de candidature : 15/05/2013
Laboratoire d’accueil : XLIM/SIC

Présentation de l’équipe de recherche
ICONES team inside XLIM-SIC department is dedicated to colour and spectral image analysis upon mathematical, perceptual or physical models.

Résumé de la thèse en français
Nous avons développé un nouveau formalisme d’analyse des images couleurs, exact théoriquement selon les modèles de la vision humaine. De ce formalisme découle des opérateurs d’extraction de contours, des attributs de textures, etc… Les fonctions de mesure de distance ou de similarité sont au coeur de ce processus.
L’objectif de ce travail est d’étendre ce formalisme aux images spectrales pour disposer de méthodes d’analyse pleines bandes (sans réduction préliminaire de la complexité).
La fonction de distance spectrale la plus adaptée est recherchée, elle devra respecter les propriétés physiques des images spectrales. Dans un second temps, nous travaillerons sur la définition et la validation d’un ensemble complet d’opérateurs de traitement d’images spectrales correct sur le plan de la physique et validés expérimentalement.

Résumé de la thèse en anglais
We developed a new formalism for colour images processing theoretically correct upon the human visual system. Thanks to it perceptual gradients, textures features are produced. Colour distances or similarity functions are the fundamentals of this formalism, especially those standardized by the CIE.
Objectives of this work are to extend this formalism to spectral images for full bands image processing (without preliminary reduction of complexity).
The most suitable distance function for spectral images is searched, this distance function will be correct upon the physical properties of spectral images. In a second step, we will focus on the definition and validation of a complete set of spectral image processing operators, theoretically correct in terms of physics and experimentally validated.

Description complète du sujet de thèse
As in the first colour image processes, the first multispectral image processing tools tried to reduce the complexity in one band or to work band by band in marginal way. Then several years will be probably required to produce vectorial approaches enabled in a perceptual way or a physical way.

Actually, the great majority of spectral processing works on reduced data, trying to reduce the complexity to one local band, to some distributions after a classification step or to three-colour band in perceptual spaces for human assessments. If such developments could be adapted for some segmentation or classification purposes, they are not adapted for textures analysis or for linear content evolution measurements. Evidently, new processing tools are required to enhance this kind of rich information and justifying acquisition with more than 3 bands.

Recently, new perspectives appear in colour image processing, in particular Richard and Ledoux have shown how to construct mathematical morphological tools from distances based functions. Thanks to this construction, for the first time, a generic framework for n-dimensional non-linear processing has been created, introducing the ability to define n-dimensional textures features or spatio-spectral objects detection.

However, behind this mathematical framework, some physical or perceptual questions are asked. In particular, the question of the right distance functions to use. As multispectral or hyperspectral data could be consider as distribution, all the mathematical background associated to this hypothesis could be used. Nevertheless, as we show actually on current works, to be adapted the distances functions must respected some constraints on linear behaviour. In addition, any classical form for distributions metrics allows these constraints.

Therefore, a dedicated job must be developed, to identify or construct the right spectral distances, in front of the different existing possibilities to measures distances or similarities. This part will develop and study the different constraints to take into account in this analysis. In particular, spectral data deals with energy distributed on frequency band, so this physical point of view induced some dedicated constraints.

As the applications domain is linked to metrological purposes or multimedia uses, the method accuracy must be expressed at the different processing level: from low-level to high level of processing. The low-level deals with spectrum ordering (in a similar way than those develop for the perceptual spaces construction) and basic morphological operators. The next levels will work on spectral textures evolution measures. Two kinds of context will be considered: the physical one and the perceptual one.

In parallel to these theoretical developments, dedicated experiments will be developed to produce the required datasets from calibrated acquisition with controlled colour textures. Identifying the right textures and controlled parameters is one of the objectives of this task to construct an image database for comparison and international benchmark on these questions.

Objectifs scientifiques de la thèse
• Define and Produce distance metrics between energy spectra, correct according to physical principles and numerically validated on reference images database.

• Construct a set of spectral image processing operators, correct according to physical principles and numerically validated on reference images database

• Produce a spectral texture image database (acquired and synthesis images) to assess spectral images processing tools. This images database will be associated to ground truth: physical reality and psychovisual evaluation for correlation with the Human Visual System.

• Define comparison criteria for spectral distances. These criteria may be low level on selectivity or linearity behaviour, or high-level on accuracy of image processing applied to the reference images database.

• Apply image processing results on real images from Cultural Heritage Domain, extract and measure elements of interest for scientist in history of Art.

Compétences à l’issue de la thèse
• Be able to design and develop an image processing chain for colour and spectral images with perceptual or physical metrological objectives.

• Be able to specify elements of decision for choosing or designing tools for spectral image processing.

• Be able to acquire colour and spectral images in controlled and calibrated environments.

• Understanding limits and perspectives for spectral and colour image processing with metrological purposes.

Mots clés (séparés par des virgules)
image processing, colour, spectral images, distance, perceptual model, physical model, acquisition, mathematical morphology
Conditions restrictive de candidature (nationalité, âge, …) : NON

Expérience/profil souhaité(e)
– Training in Image processing.

– Knowledge in colour image processing domain and colour image acquisition are searched.

Modalité de dépôt de candidature
Send an email:
– detailing your previous activity inside the colour image processing
– with a CV
– and some elements justifying your rank in your last learning steps.
Envoyer CV, avec relevé de notes M1 et M2, et lettre de motivation au directeur de thèse.

Date limite de candidature
30 avril 2013

Directeur de thèse
Noël RICHARD, Christine Fernandez-Maloigne
Adresse mail du directeur de thèse : richard@sic.univ-poitiers.fr
Téléphone Directeur de thèse : 33 5 49 49 66 20

Co-directeur de thèse
Jon Yngve Hardeberg, PhD, Professor of Color Imaging
The Norwegian Colour and Visual Computing Laboratory
Faculty of Computer Science and Media Technology
Gjøvik University College, PO Box 191, N-2802 Gjøvik, Norway
Adresse mail du co-directeur de thèse : jon.hardeberg@hig.no
Cofinancement LABEX SigmaLIM demandé : NON

Recherche

Menu principal

Haut de page