A new procedure for the modelling and representation of classes in multivariate images

Abstract

This paper describes a new procedure for the estimation of classes in multivariate images. The Feedback Multivariate Model Selection (FEMOS) procedure combines unsupervised and supervised classifiers with a model evaluation criterion to extract classes from the multivariate image in an iterative manner. The procedure uses a subset of the multivariate image to estimate a general model and evaluates this model with the original multivariate image via a model evaluation criterion. The procedure can be applied for the unsupervised segmentation of multivariate images or for training and test set estimation from multivariate images. Furthermore, a new coloring scheme, called class coloring, is presented for coloring of class labels in segmented images. The coloring of class labels is automated, which makes it independent of the number of classes and shows more resemblance with the pseudocolor multivariate image. The procedure is tested on different real world multivariate images and the results are compared with cluster size-insensitive FCM (csi-FCM), a clustering algorithm. The results show that the procedure outperforms traditional routines in terms of robustness and accuracy when applied as either unsupervised segmentation technique or class modelling technique.