A low-rank tensor-based algorithm for face recognition

摘要

The face recognition problem arises in a wide range of real life applications. Our new developed face recognition algorithm, based on higher order singular value decomposition (HOSVD) makes use of only third order tensor. A convenient way of writing the commuta-tivity of different modes of tensor-matrix multiplications leads to a new method that outperforms in terms of complexity another third order tensor method. The resulting algorithm is more successful (in terms of recognition rate) than the conventional eigenfac-es algorithm. Its effectiveness is proved for two benchmark datasets (ExtYaleB and Essex datasets), which are ensembles of facial images that combine different modes, like facial geometries, illuminations, and expressions.