Naive Bayes classifier is a simple and effective classification method, but its attribute independence assumption makes it unable to express the dependence among attributes, and affects its classification performance. On the basis of analyzing the classification principle of Bayesian classification model and a variant of Bayes theorem, a new classification model based on Bayes theorem, DLBAN (double-level Bayesian network augmented naive Bayes), which adds the dependence among attributes by selecting the key attributes, is proposed. DLBAN classifier is compared with Naive Bayes classifier and TAN (tree augmented naive Bayes) classifier by an experiment. Experimental results show this model has higher classification accuracy in most data sets.