Texture classification with fisher kernel extracted from the continuous models of RBM

In this paper, we introduce a novel technique of deriving Fisher kernels from the Gaussian Bernoulli restricted Boltzmann machine (GBRBM) and factored 3-way restricted Boltzmann machine (FRBM) to yield better texture classification results. GBRBM and FRBM, both, are stochastic probabilistic models that have already shown their suitability for modelling real valued continuous data, however, they are not efficient models for classification based on their likelihood performances (Jaakkola and Haussler, 1999; Azim and Niranjan, 2013). We induce discrimination in these models with the help of Fisher kernel that is constructed from the gradients of the parameters of the generative model. From the empirical results shown on two different texture data sets, i.e. Emphysema and Brodatz, we demonstrate how a useful texture classifier could be built from a very compact generative model that represents the data in the Fisher score space discriminately. The proposed discriminative technique allows us to achieve competitive classification performance on texture data sets, without expanding the size of the generative model with large number of hidden units. Also, comparative analysis shows that factored 3-way RBM is a good representative model of textures, giving rise to a Fisher score space that is less sparse and efficient for classification.