Facial Kinship Verification using Tensor Discriminative Subspace based Color Components

In this work, we propose the evaluation of color spaces of 2D facial images in order to determine the most discriminating color component for kinship facial verification in uncontrolled environments. Using multidimensional facial data modeling, our approach is based on higher order tensors. Multilinear sub-space projection is provided by a new method called Multilinear Side-Information Based Discriminant Analysis (MSIDA). Several inter-connected subspaces are obtained on different tensor modes, so that the subspaces are learned iteratively by unfolding the tensor in the different modes. Using only weak information, MSIDA projects the face tensor into a new subspace in which the discrimination is improved. The input of our system is one of the color spaces RGB, YcbCr and HSV,in which, each component is evaluated separately. The experimental evaluation in the Cornell KinFace database shows that the proposed approach outperforms significantly the most state-of-the-art approaches.