Object recognition with the HOSVD of the multi-model space-variant pattern tensors

Abstract

The paper presents a framework for object recognition with the multi- model space-variant approach in the log-polar domain built into the multilinear tensor classifier. Thanks to this the method allows recognition of rotated and/or scaled objects taking advantage of the foveal and peripheral information. Recognition is done in the multilinear subspaces obtained after the higher-order singular value decomposition of the pattern tensor. The experiments show high accuracy and robustness of the proposed method. © 2011 Springer-Verlag.