Skip to Main content Skip to Navigation
Theses

Optimisation hybride mono et multi-objectifs de modèles actifs d'apparence 2,5D pour l'analyse de visage

Abstract : In this study we are interested in the pose estimation and precise localization of face features such as the eyes, the nose and the mouth of an out-of-plane rotated unknown face. Main application of this thesis work is in the Cognitive Radio equipments. We place ourselves within the framework of a low quality acquisition with camera(s) installed on Cognitive Radio equipments e.g. mobile phone, laptop, desktop computers etc. The face pose and localization of facial features in an unconstrained environment are the major problems for CR equipment. All of its subsequent face related applications (e.g. face recognition, face synthesis, face data compression etc.) highly depend upon the methods used for the facial analysis system. In order to extract face features, we use the Active Appearance Models (AAM), deformable models allowing shape and texture to be jointly synthesized. We initially propose a new 2.5D AAM, based on 3D model, which makes it possible to perform pose estimation and features localization of an oriented face. Secondly, we propose a new optimization methodology for the face search by AAM, in a single camera system, by the hybridization of deterministic and direct search method which has never been used and tested before. Our method hybridizes Gradient Descent (GD) inside the Genetic Algorithm (GA) in a unique way. Along with other operators of GA we propose gradient operator which works in conjunction with the mutation operator of GA thus it does not make the system computationally expensive. Finally for a complete facial analysis system by multiple cameras, we proposed a new concept of multi-objective AAM. In this method, facial images from multiple cameras are analyzed simultaneously by 2.5D AAM. For the face search optimization we propose a unique way of hybridizing GD with NSGA-II (Non-dominating Search Genetic Algorithm-II). Both of our propositions are robust, real time, efficient and extract facial features even in unknown and out-of-plane rotated faces.
Complete list of metadata

Cited literature [127 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00491328
Contributor : Myriam Andrieux Connect in order to contact the contributor
Submitted on : Friday, June 11, 2010 - 11:34:32 AM
Last modification on : Wednesday, April 27, 2022 - 3:54:51 AM
Long-term archiving on: : Friday, September 17, 2010 - 1:50:06 PM

Identifiers

  • HAL Id : tel-00491328, version 1

Citation

Abdul Sattar. Optimisation hybride mono et multi-objectifs de modèles actifs d'apparence 2,5D pour l'analyse de visage. Informatique [cs]. Université Rennes 1, 2010. Français. ⟨tel-00491328⟩

Share

Metrics

Record views

462

Files downloads

160