Automatic facial expression recognition is useful to create applications in various domains such as health, road safety or marketing where feedback on user state is relevant. Despite very good results in controlled settings (frontal face, no occlusion, goodillumination), facial expression recognition is still today challenging under unconstrained environment. Among the different challenges, occlusions are particularly difficult to handle as they add noisy elements on the images and hide parts of the information. Several solutions have been proposed to address this issue. These solutions can be roughly categorized in two : those which focus on visible regions of the face and those which try to reconstruct the hidden part. State-of-the-art solutions are mainly based on texture or, sometimes, geometry and few are based on movement. However, movement seems to be particularly adapted under occlusions thanks to different motion properties such as close range propagation and local coherency. This work show the interest of using movement to overcome the issue of occlusions for the task of facial expression recognition.
- Directeur de thèse : Chaabane Djeraba - Rapporteurs : Frédéric Jurie, Yannick Benezeth - Examinateurs : Saïda Bouakaz, Laetitia Jourdan, Bruno Bachimont, Hichem Sahbi, Nacim Ihaddadene, Matthieu Cord (invité), Ioan Marius Bilasco (invité)
Thesis of the team FOX defended on 01/09/2021
français
English