| Texture classification of radiographic images using Markov random field (MRF) modeling in a bayesian and decision theory framework is addressed. Images are viewed as being formed by a hierarchical structure which may include various texture definitions. Two general classification procedures are presented, both based on the estimation of the a posteriori probabilities for texture class. One procedure estimates the a posteriori distribution of texture class and involves Gibbs sampling and Monte Carlo integration. The other method provides an estimate of the maximum a posteriori (MAP) texture class by combining Gibbs sampling and simulated annealing. The parameterization of the associated Gibbs probability measures is performed using a simple histogramming approach. Both classification methods are tested on selected regions from actual radiographic images and simulated images for several signal to noise ratio (SNR) conditions. As an upper bound on classification performance, signal known exactly (SKE) cases are presented for each test image. Results show very good classification performance for all cases studied. Furthermore, and since classification and not enhancement is sought, they show that both algorithms converge rapidly towards the correct classification, reducing the number of sampling steps and simplifying the overall computational complexity. |
