Abstract: Feature level fusion plays an important role in the process of data fusion. Acco rding to the idea of canonical correlation analysis(CCA), a novel method of comb ined feature extraction is proposed in this paper. The main idea of this method can be described as follows. First of all, two groups of feature vectors with th e same pattern sample are extracted, and the correlation criterion function betw een the two groups of feature vectors is established. Then, based on this criter ion function, their canonical correlation features are extracted to form effecti ve discriminant vectors for classification. The advantage of this method lies in the following aspects: firstly, it suits for information fusion; secondly, it e liminates the redundant information within the features, and a new way for class ification and recognition is proposed. In addition, the essence of the efficienc y is analyzed further in theory. The results of experiments on Yale and ORL face databases show that the recognition rate is far higher than that of the recogni tion adopting the single features, and that this algorithm is efficient and robu st.