Analyse d'expressions faciales par modèles d'apparence
Bouchra Abboud · 2004
Verbal expression fluency and rhetorical ease are incontestable aspects of successful communication. However, humans are able to communicate in a variety of ways besides the use of words, including gestures and facial expressions. As a matter of fact the idiom poker face evokes an attitude of blank expression to prevent detection of intent which suggests that facial expressions constitute an essential modality in human communication. This thesis addresses the issue of representations for facial expression recognition and synthesis. Ln this context, a global appearance model is used in conjunction with bilinear factorization allowing to separate expression specifie factors from identity specific factors in the global appearance parameters. A feature extraction technique inspired from the above representations is then proposed which consists in automatically computing the optimal identity and expression components that best adapt to an unknown target face. Facial expression recognition and synthesis are finally performed using each representation and their performances are compared quantitatively and qualitatively. Factorization based models yield very interesting synthesis performances in terms of visual quality of the synthetic faces.