Evaluation methodology for classification process of digital images

A methodology is proposed for selecting the best training set, the best classifier and the best set of parameters associated to the classifier. A performance indicator is defined from measures of signatures separability, accuracy between signatures in the training set and in the classified final image, and probability that a pixel belongs to the class which has been assigned. The theoretical description of the suggested methodology has been proved using a image from the Fourier time series analysis. The algorithm has been executed and the obtained results for the best and the worst classifier have been analyzed and evaluated.