Classification of textures using noncausal hidden Markov models

Texture classification is a two step process. The first consists of generating a model of each texture type using any of a number of texture modeling techniques. The second step consists of classifying a given texture into a category by comparing it against the training models and choosing the category or model which is closest in some sense to the texture being classified. This dissertation is concerned with the use of noncausal hidden Markov models (HMMs) for texture classification. Hidden Markov models have been successfully used in speech processing and have recently been used in image processing applications. The HMM assumes that an image can be modeled by a statistical process whose states are not directly observable and that each state has a statistical distribution of an observable quantity such as reflectance (gray level).; HMMs may be causal or noncausal. For example, causality may be the assumption that the state of each pixel is dependent on the state of neighbors above and to the left of it. In noncausal models the state of each pixel may be dependent on its neighbors in all directions. In this dissertation textures are modeled principally by noncausal HMMs. New algorithms are developed to learn the parameters of the HMM of a texture and to classify it into one of several learned categories. Learning and classification algorithms were developed for four and eight neighbor noncausal models and a three neighbor causal model. In addition, both absolute and ranked gray levels are used as observations.; The effectiveness of these algorithms in texture classification has been determined by use of many classification experiments involving both synthetically generated and natural textures. The classification accuracies obtained using the noncausal models are compared to those obtained using the three neighbor causal model. As a control, the results are also compared to results obtained on the same textures with an autocorrelation based texture classification algorithm. Finally, some additional results are shown where state determination is based on texture orientation rather than gray level but with absolute gray levels used as observations.