Classification of scenes based on multiway feature extraction

Recognition of real world scenes can be efficiently solved based on global features termed the Spatial Envelope. Such features indeed comprise multiple modes such as orientations, scales, sparsity profiles. In order to extract features and classify multiway samples, most approaches vectorize data tensors to convert the classification of multiway data into the one of 1-D samples. This common approach disregards the multiway structures of global features, hence it can face the risk of losing correlation information between modes (orientations or scales). To this end, by revisiting the problem of scene classification in view of tensor decompositions, a new method is introduced to extract multiway features. The projection filter is designed for global features based on a set of basis matrices instead of only one basis as in 1-D problem. The proposed approach not only improves the classification accuracy, but also reduces the running time for training stage and feature projection.