Textured image recognition using hidden Markov model

It is shown that when textures are modeled as Markov chains, the per symbol entropy of a 1-D texture profile can be used as a classification criterion. both noiseless and noisy test textures are studied, and five methods of classification are developed, based on whether and how the knowledge of noise distribution is given. For six random microtextures, an 80-90% correct classification rate is achieved for moderate to low noise power levels. This suggests that much 2-D textural information is preserved in a 1-D profile when a Markov chain is used to model textures