Multidimensional Image Classification based on Influence Controllers of the Classes in the Images

The Bayesian statistical approach is a well known technique used in computer based image classification, and the Maximum Likelihood Classifier (ML) is one of the most present in literature. The structure of the ML Classifier is such that, in a multidimensional approach, every image used in the classification process has the same effect or contribution, regardless its intrinsic quality. This paper extends the concept of image influence controller in the Modified Global Membership Function (MGMF) to classes reliability factors in order to use class influence controllers instead a single image reliability factor. It is also proposed an estimator for the classes reliability factors based on the Confusion Matrix and its Conditional Kappa coefficient. Two SAR images are used to evaluate the estimator and the classification process that take into account the classes reliability factors.