Multilinear Principal Component Analysis Network for Tensor Object Classification

The recently proposed principal component analysis network (PCANet) has performed well with respect to the classification of 2-D images. However, feature extraction may perform less well when dealing with multi-dimensional images, since the spatial relationships within the structures of the images are not fully utilized. In this paper, we develop a multilinear principal component analysis network (MPCANet), which is a tensor extension of PCANet, to extract the high-level semantic features from multi-dimensional images. The extracted features largely minimize the intraclass invariance of tensor objects by making efficient use of spatial relationships within multi-dimensional images. The proposed MPCANet outperforms traditional methods on a benchmark composed of three data sets, including the UCF sports action database, the UCF11 database, and a medical image database. It is shown that even a simple one-layer MPCANet may outperform a two-layer PCANet.