A three stage scheme for classification of multi-spectral images is proposed. In each stage, statistics of each class present in the image are estimated. The user is required to provide only one prototype pixel for each class to be seeded into a homogeneous region. The algorithm starts by generating optimum initial training sets, one for each class, maximizing the redundancy in the data sets. These sets are the realizations of the maximal discs centered on the prototype pixels for which it is true that all the elements belong to the same class as the center one. Afterwards a region growing algorithm increases the sample size providing more statistically valid samples of the classes. Final classification of each pixel is done by comparison of the statistical behavior of the neighborhood of each pixel with the statistical behavior of the classes. A critical sample size obtained from a model constructed with experimental data is used in this stage. The algorithm was tested with the Kappa coefficient k on synthetical images and compared with K-means (k~=0.41) and a similar scheme that uses spectral means (k~=0.75) instead of histograms (k~=0.90). Results are shown on a dermatological image with a malignant melanoma.