Focusing on the problem that the process of removing redundancy will be very complicated due to the large number of the selected features, and the problem that some features only can have strong correlation with label after being combined with other features in the feature selection, a Multi-Level Feature Selection algorithm based on Mutual Information (MI_MLFS) was proposed. Firstly, the features were divided into strongly correlated, sub-strongly correlated and other features according to the degrees of correlations between features and label. Secondly, after selecting strongly correlated features, features with low redundancy in the sub-strongly correlated features were selected. Finally, the features which were able to enhance the correlation between the selected feature subset and label were selected. Among 15 datasets, MI_MLFS was compared with the algorithms of ReliefF, minimal-Redundancy-Maximal-Relevance criterion (mRMR), Joint Mutual Information (JMI), Conditional Mutual Information Maximization criterion (CMIM) and Double Input Symmetrical Relevance (DISR). The results show that MI_MLFS achieves the highest classification accuracy in 13 datasets and 11 datasets with Support Vector Machine (SVM) classifier and Classification And Regression Tree (CART) classifier respectively. MI_MLFS has better classification performance than many classical feature selection algorithms.