Multilinear Enhanced Fisher Discriminant Analysis for robust multimodal 2D and 3D face verification

摘要

In the last few years, multilinear subspace learning has attracted large interest for dimensionality reduction and classification of multidimentional data. In this paper, we propose a new multilinear supervised projection method, called Multilinear Enhanced Fisher Discriminant Analysis (MEFDA), which improves the generalization capacity by decomposing the discriminate analysis process into a simultaneous diagonalization of the inter-class and intra-class scatter matrices resulting from the unfolding of tensor data. MEFDA uses the discriminant tensor criterion based on iterative procedure in order to obtain the optimal projection matrix in each k mode of tensor data. One virtue of MEFDA is that it can avoid the disadvantage of large dimensionality, in which the computational cost is reduced. Since the data encoded using high-order tensor, MEFDA respects the original structure and the natural geometry of face data. Therefore, the natural correlation of multi-dimension face data can be exploited. The proposed algorithm has been evaluated on three public 3D face datasets, FRGC V2.0, CASIA 3D and Bosphorus under the hard challenges of poses, expressions, illuminations and occlusions. The experimental results demonstrate that the proposed method significantly outperforms the current state of the art methods, with 99% and 97.02% at 0.001 False Acceptance Rate (FAR) for FRGC v2.0, and Bosphorus datasets, respectively, and 99.36% at evaluation Equal Error Rate (EER) for CASIA 3D dataset.