Face Recognition using Tensors of Census Transform Histograms from Gaussian Features Maps

Face recognition is a challenging task due to the large variety of appearance that a face may exhibit under variations in illumination and viewing position as well as variations in facial expression. Many of the more successful approaches use Gabor wavelets as an image descriptor, resulting in relatively high computational cost. In this work, we have explored the use of Gaussian derivative features calculated with a linear complexity half-octave Gaussian pyramid [2]. We propose a tensorial representation for Gaussian derivative histograms that associate Gaussian features by the nature of information that is encoded. This representation retains spatial structure without loss of information due to vectorization of features. The over-all architecture of our face recognition system is showed in the Figure 1 (B). We can summarize our method as follows: first, local differences in the half-octave Gaussian pyramid are used to compute Gaussian feature maps. For each location in each Gaussian map, a Census Transform histogram is calculated and concatenated to form a local tensor. Only one tensor is computed for each class of Gaussian map. An MPCA (Multilinear Principal Component Analysis) [3] method is then applied to each tensor to reduce the dimensionality and correlation due of Census Transform. Finally we apply the KDCV (Kernel Discriminative Common Vectors) method [1] to generate a discriminative vector.