Auteur / Autrice : | Noura Azzabou |
Direction : | Nikos Paragios, Frédéric Guichard |
Type : | Thèse de doctorat |
Discipline(s) : | Mathématiques. Informatique |
Date : | Soutenance en 2008 |
Etablissement(s) : | Paris Est |
Mots clés
Mots clés contrôlés
Mots clés libres
Résumé
This thesis is devoted to image enhancement and texture preservation issues. This task involves an image model that describes the characteristics of the recovered signal. Such a model is based on the definition of the pixels interaction that is often characterized by two aspects (i) the photometric similarity between pixels (ii) the spatial distance between them that can be compared to a given scale. The first part of the thesis, introduces novel non-parametric image models towards more appropriate and adaptive image description using variable bandwidth approximations driven from a soft classification in the image. The second part introduces alternative means to model observations dependencies from geometric point of view. This is done through statistical modeling of co-occurrence between observations and the use of multiple hypotheses testing and particle filters. The last part is devoted to novel adaptive means for spatial bandwidth selection and more efficient tools to capture photometric relationships between observations. The thesis concludes with providing other application fields of the last technique towards proving its flexibility toward various problem requirements.