Canonical Correlation Analysis (CCA) is a powerful multi-mode feature fusion method, but in traditional CCA, the optimization function is to find a pair of projections which make the mappings of the observations of the same pattern have the maximum correlation. It is an unsupervised subspace learning algorithm. This paper, we propose a Improved Canonical Correlation Analysis (ICCA) method which improves the optimization function by adopting the supervised relationships of the patterns belong to the same class. Our proposed algorithm is validated by the experiments on Jaffe facial expression database. Be compare with other methods, the recognition rate of our method is far higher than the PCA algorithm which only adopts single-mode image features and the traditional multi-mode CCA feature fusion algorithm.