Estimation of class correlation parameters in images for context classification

A wide class of models for the probabilistic dependency of class labels in a neighborhood within an image is defined. A convergent and computationally efficient closed-form estimator for these context-defining parameters is developed. The estimation procedure assumes that the class conditional densities are known and operates on unlabeled image data. The estimator can be recursively implemented. Examples of practical context models within the class of models are given. The appropriate model can also be chosen with the aid of the context parameter estimates. Applications of the estimator in existing context classifiers are pointed out