UNSUPERVISED CONTEXT ESTIMATION IN A MESH OF PATTERN CLASSES FOR IMAGE RECOGNITION

This paper defines a wide class of models for the statistical dependency of class labels (context) in a neighborhood within an image. A computationally efficient closed form estimator for these context defining parameters is developed. The estimator is unbiased, convergent in the mean square sense, and can be recursively implemented. Parameters of several practical context models within the class of models are estimated as functions of the estimates of the parameters in the general model. The appropriate model can also be decided with the help of the context parameter estimates. Applications of the estimator in existing context classifiers are mentioned. A limitation of the estimator is also pointed out. Iterative approaches for decision making in Markov models are outlined to highlight the use of our estimator to ensure the convergence of iterative decisions to a local optimum.