The automatic classification of objects is a very interesting approach under several problem domains. This paper outlines some results obtained under different classification models to categorize textural patterns of minerals using real digital images. The data set used was characterized by a small size and noise presence. The implemented models were the Bayesian classifier, Neural Network (2-5-1), Support Vector Machine, decision tree and 3-nearest neighbours. The results after applying crossed validation show that the Bayesian model (84%) proved better predictive capacity than the others, mainly due to its noise robustness behaviour. The Neuronal Network (68%) and the SVM (67%) gave promising results, because they could be improved increasing the data amount used, while the decision tree (55%) and k-NN (54%) did not seem to be adequate for this problem, because of their sensibility to noise.