In terms of the distribution of objects and their classification labels, the continuous feature space was partitioned into a variety of subspaces, each one with clear edge and unique classification label. After the projection of all the subspaces to each feature, the quality of each feature was estimated for a subspace opposite to all the other subspaces with different classification labels by means of statistical significance. Through construction of a matrix by all the estimate qualities of all features of all the subspaces, all the features was ranked from the highest classifying power to the lowset on the matrix for the feature space. According to the ranked-feature set, the feature selection was completed. The experimental results illustrate that the feature selection is efficient and effective.