Combination of projectors, standard texture descriptors and bag of features for classifying images

In this work we first propose a new approach to improve the information that can be extracted from the texture descriptors used for describing a given image. These descriptors are used to study the usefulness to combine a classifier trained with "standard" application of texture descriptors (i.e. a global extraction of descriptors from the whole image or a set of descriptors locally extracted from each sub-window of the image and then concatenated) and a bag-of-features approach. In our bag-of-features approach a set of features is extracted from each sub-window of the image, these sets are quantized, and the resulting global descriptor vectors are used to train a support vector machine based classifier. For improving the information that could be extracted from an image we propose to use projectors. A projector (Pi) maps a vector y to the subspace defined by (Pi) and a space of orthogonal projectors induce a probability distribution for every y. Since it is difficult to describe each class with only one projector we extract different clusters from each class and from each cluster a different projection space is found. The Matlab code for the bag of feature approach will be publicly available at https://www.dei.unipd. it/node/2357 (C) 2015 Elsevier B.V. All rights reserved.