Principal component analysis(PCA) is a common method for feature selection. The classical procedure to obtain principal components is calculating the correlation matrix between features. However, the correlation cannot reflect the nonlinear relationship. Mutual information measures the interdependence strength between variables which are not limited to the linear correlation. PCA based on mutual information(MIPCA) for feature selection is presented. The algorithm calculates the mutual information matrix and extracts the eigenvalues as the criteria to determine the number of principal components and assess the effect of feature selection. Finally, the proposed algorithm is compared with PCA by cases, and the efficiency of classification is tested by neuron network.