3D Face Recognition based on Pose and Expression Invariant Alignment

Abstract

3D face recognition has made considerable progress during the last decade as an emerging biometric modality. In order to ensure reliable 3D face recognition, novel 3D alignment and recognition algorithms are proposed in this research work. The principal objective of this dissertation is to investigate and introduce novel techniques to construct a fully automatic 3D facial recognition system. The first study presents a novel, pose and expression invariant approach for 3D face alignment based on intrinsic coordinate system (ICS) characterized by nose tip, horizontal nose plane and vertical symmetry plane of the face. It is observed that distance of nose tip from 3D scanner is reduced after pose correction which is presented as a quantifying heuristic for the proposed alignment scheme. In addition, motivated by the fact that a single classifier cannot be generally efficient against all face regions, a two tier ensemble classifier based 3D face recognition approach is presented which employs Principal Component Analysis (PCA) for feature extraction. The individual regions are classified using Mahalanobis Cosine (MahCos) distance, Euclidean distance, Mahalanobis (Mah) distance, and Manhattan distance in separate experiments. The resulting matching scores are combined using weighted Borda Count (WBC) based combination and a re-ranking stage. The performance of the proposed approach is corroborated by extensive experiments performed on two databases, namely, FRGC v2.0 and GavabDB, confirming effectiveness of fusion strategies to improve performance. In the second study, a novel and fully automatic pose and expression invariant 3D face recognition algorithm is proposed using two-pass 3D face alignment based on minimum distance and two-pass 3D face alignment based on classification approach. The proposed alignment approaches are capable of aligning neutral and expressive 3D faces acquired at frontal and non-frontal poses whereas the former is capable of aligning profile face images as well. For the face recognition framework, multi-view 3D faces are synthesized to exploit real 3D facial information. The matching scores are computed between multi-view face images using Mahalanobis Cosine (MahCos) viii distance, Euclidean distance, Mahalanobis (Mah) distance and Manhattan distance in separate experiments. Inspired by the effectiveness of fusion approaches, Support Vector Machine (SVM) is employed using scores obtained from multi-view face pairs for face verification. In addition, a three stage unified classifier based face identification algorithm is employed which combines results from seven base classifiers at first stage, two parallel face recognition algorithms at second stage and an exponential rank combiner at third stage in a hierarchical manner. For profile face images, the face identification algorithm combines results using four base classifiers, two parallel face recognition algorithms and the rank combiner stage. The performance of the proposed methodology is demonstrated by extensive experiments performed on two databases: FRGC v2.0 and GavabDB. The results show that the proposed methodology can be efficiently used to construct a pose and expression invariant facial recognition system.