Multimodal 2d+3d multi-descriptor tensor for face verification
Résumé
In the last few years, there is a growing interest in multilinear subspace learning for dimensionality reduction of multidimensional data. In this paper, we proposed a multimodal 2D + 3D face verification system based on Multilinear Discriminant Analysis MDA integrating Within Class Covariance Normalization WCCN technique. Histograms of local descriptor applied to features extraction from 2D and 3D face images are concatenated and organized as a tensor design. This tensor is then reduced and projected using MDA technique into a lower subspace. WCCN technique is used to reduce the effect of the intra class directions using normalisation transform and to enhance the discrimination power of the MDA. Our experiments were carried out on the three biggest databases: FRGC v2.0, Bosphorus and CASIA 3D under expressions, occlusions and pose variations. Experimental results showed the superiority of the proposed approach in term of verification rate when compared to the state of the art method.