Pose invariant face recognition: 3D model from single photo - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Optics and Lasers in Engineering Année : 2017

Pose invariant face recognition: 3D model from single photo

Thibault Napoléon
Ayman Alfalou

Résumé

Face recognition is widely studied in the literature for its possibilities in surveillance and security. In this paper, we report a novel algorithm for the identification task. This technique is based on an optimized 3D modeling allowing to reconstruct faces in different poses from a limited number of references (i.e. one image by class/person). Particularly, we propose to use an active shape model to detect a set of keypoints on the face necessary to deform our synthetic model with our optimized finite element method. Indeed, in order to improve our deformation, we propose a regularization by distances on graph. To perform the identification we use the VanderLugt correlator well know to effectively address this task. On the other hand we add a difference of Gaussian filtering step to highlight the edges and a description step based on the local binary patterns. The experiments are performed on the PHPID database enhanced with our 3D reconstructed faces of each person with an azimuth and an elevation ranging from −30° to +30°. The obtained results prove the robustness of our new method with 88.76% of good identification when the classic 2D approach (based on the VLC) obtains just 44.97%.
Fichier non déposé

Dates et versions

hal-02562739 , version 1 (04-05-2020)

Identifiants

Citer

Thibault Napoléon, Ayman Alfalou. Pose invariant face recognition: 3D model from single photo. Optics and Lasers in Engineering, 2017, 89, pp.150-161. ⟨10.1016/j.optlaseng.2016.06.019⟩. ⟨hal-02562739⟩
66 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More