Kernel Canonical Correlation Analysis (KCCA) is a recently addressed supervised machine learning methods, which is a powerful approach of extracting nonlinear features. However, the standard KCCA algorithm may suffer from computational problem as the training set increases. To overcome the drawback, an improved KCCA was proposed. Firstly, a scheme based on geometrical consideration was proposed to select a subset of samples that were projected to feature space (Reproducing Kernel Hilbert Space). And then, an efficient algorithm was proposed to enhance the efficiency of the feature extraction, which selected the most contributive eigenvectors for training and classification, and then calculated the corresponding eigenvectors for classification. Finally, the improved KCCA was combined with a multi-class classification method based on Support Vectors Data Description (SVDD) for classification and recognition, which put forward new ideas for pattern recognition based on kernels. The experimental results on ORL face database show that the proposed method reduces the training time and the system storage without deteriorating the recognition accuracy compared with standard KCCA.