Texture based classification of images using frequency estimated pairwise MRF joint distributions on site labels from wavelet decomposed images

In this paper we demonstrate the efficacy of using joint probabilities on the values (pixel intensities/wavelet coefficients) for neighbouring sites (pixels/spatially neighbouring wavelet coefficients), to classify images based on texture. The classification capacity for this type of joint distribution, used as a feature, is tested using a first nearest neighbour (NN1) method, which counts the number of errors when comparing the labeled texture to the label calculated by assigning the texture to the class of its nearest neighbour, calculated using our method. We compare our classification results to another method based on histogram comparison. Our classification methodology is simple, general, extensible and fast to calculate.