Tensor-Jet: A tensorial representation of Local Binary Gaussian Jet maps

In this paper we present a new robust method for recognizing face images using a robust tensorial representation of binary gaussian jet maps (Tensor-Jet). This tensorial representation captures local appearance while retaining information about the spatial structure. During the tensors construction, each Gaussian Jet map is calculated with a Half Octave Gaussian Pyramid using a linear complexity algorithm. A Local Binary Pattern (LBP) operator is then applied and the results are accumulated in a local histogram. Local histograms are concatenated to form a tensorial representation that captures the spatial structure. The local correlation of neighboring histogram is removed by applying multi-linear principal components analysis. Finally a Kernel Discriminative Common vector is trained with the output of the MPCA to improve the overall recognition. We compare two different algorithms to recognize face images with this representation. Experimental results using the FERET database and Extended Yale Database show that this method compares favorably with the state-of-the-art methods in face recognition.