Bayesian texture classification method using a random sampling scheme

We present a texture classification approach that uses a Bayesian inference procedure using local co-occurrence properties over a set of randomly sampled points as evidence. Prior probabilities are modelled using gray level co-occurrence matrices (GLCMs) in a number of distances and orientations. By using Bayes´ rule, we find texture class that maximizes a posteriori probability of the observed gray level intensity pair in a randomly chosen point. Each point casts a vote for the texture that best explains observed co-occurrence properties. A majority voting procedure assigns a winning label for a texture class. Our approach results in a fast classifier because it does not need to compute GLCM for the texture under test. Our method was tested on a subset of textures from Brodatz database and the classifier accuracy was estimated at about 85% even when a small fraction of points in the image under test were used for the classification phase.