A tensor-based multivariate statistical model for 3D face and facial expression recognition

A major difficulty in face recognition is caused by variation of pose and lighting in the image acquisition process, and expression differences in the subject. Three dimensional methods are inherently invariant to pose, and to some degree to lighting as well, and so have proved more accurate than their two dimensional counterparts. However they are still susceptable to errors caused by expression variation. We have investigated a technique to explicitly model facial expressions using subspaces in a multivariate statistical discriminant tensor model. We use multiple independent lower dimensional discriminative subspaces to extract facial features. The framework has the ability to characterise face variation across subjects, and therefore can deal with the shortcomings of conventional statistical discriminant methods. Recognition experiments were conducted using 3D face database created by the State University of New York (SUNY), Binghamton. The results show higher face and facial expression recognition rates compared with the more popular eigenface techniques.