Gated Boltzmann machine in texture modeling

Abstract

Recently, a new type of neural network method, namely deep learning, was discovered, which yielded excellent results in various tasks such as handwritten digit recognition, speech recognition and facial expression recognition. Different from traditional multi-layer perceptron learning methods (MLP), an unsupervised pre-training step before supervised learning is of huge importance in learning successful features. Also, it is argued that the deep architecture has more expressive power comparing to the conventional shallow networks, such as support vector machine or multi-layer perceptron's.

Even though deep learning has yielded a large amount of world-class records in different tasks, there is little research on how the deep network can be used in texture analysis.

For this particular problem, we consider the problem of modelling complex texture information using undirected probabilistic graphical models. Texture is a special type of data that one can better understand by considering its local structure. For that purpose, we propose a convolutional variant of the Gaussian gated Boltzmann machine (GGBM), inspired by the co-occurrence matrix in traditional texture analysis. We also link the proposed model to a much simpler Gaussian restricted Boltzmann machine where convolutional features are computed as a pre-processing step. The usefulness of the model is illustrated in texture classification and reconstruction experiments.