Pose - Invariant Multimodal ( 2 D + 3 D ) Face Recognition using Geodesic Distance Map

In this paper, an efficient pose-invariant face recognition method is proposed. This method is multimodal means that it uses 2D (color) and 3D (depth) information of a face for recognition. In the first step, the geodesic distances of all face points from a reference point are computed. Then, the face points are mapped from the 3D space to a new 2D space. The proposed mapping is robust under the in-depth face rotations. Finally, the feature extraction and face classification task is done in the new 2D space. For feature extraction, we use the Patch Pseudo Zernike Moments (PPZM) with a new weighting method to decline the self-occlusion caused by in-depth rotations. For this purpose, a novel approach for self-occlusion detection based on geodesic distances of face points is proposed and a self-occlusion map is created. For evaluation purpose, a large scale 3D face database is used and the various in-depth rotations (vertical and horizontal) are tested. The performance of the proposed method in two scenarios is compared with a classical 3D face recognition method. The results emphasize the performance of the proposed method in the pose-invariant face recognition. [Farshid Hajati, Abolghasem A. Raie. Pose-Invariant Multimodal (2D+3D) Face Recognition using Geodesic Distance Map. Journal of American Science 2011;7(10):583-590]. (ISSN: 1545-1003). http://www.americanscience.org.