Summary: The problems of texture images classification and feature space reduction are considered. The multialternative classification is reduced to a binary one-dimensional problem, for which it is possible to use Bayesian approach with one-dimensional estimates of distribution. Hypothesis of beta-distribution for one separate class is introduced. Distribution parameters are estimated by method of moments. To estimate four parameters, analytic expressions and statistic estimations of first four moments of this distribution are needed. After that the hypothesis of distribution is verified by Pearson/s criterion. It is established experimentally that beta-distribution model is generally applicable to estimate feature values distribution. It is concluded that such checks are needed for every learn sample. Feature effectiveness is estimates by analysis of degree of estimated distributions intersection. The cross-correlation of selected features is considered. A method for estimating the informative value of feature is introduced. It is based on the minimum average error probability for one feature and on mutual uncorrelatedness for a system of features. For each pair of classes we build feature system using the algorithm for estimation of informative value. The classifier, which uses feature systems and makes decision on the basis of density estimation by beta-distribution model on the stage of the binary problem, is built. This classifier also aggregates the final decision from the results of binary problems decisions.