Support Vector Machine-based method is pre sented to recognize human faces. Compared with traditional methods, such as Nearest Neighbor Rules, Euclidian Distance, Mahalanobis Distance and Neural Networks, this method achieves higher recognition rate. It can classify complicated patterns and overcome disadvantages of overfitting in traditional methods. The process of this methed is as follows: pre-processing the human face images first, then using primary component analysis (PCA) to extract and select the appropriate features of human faces, training multiple SVMs by the face feature vectors, and using the trained SVMs to classify human faces at last. The performance of SVMs are also compared. It is concluded by experiments that the second-order polynomial SVM has better performance than other SVMs on human face recognition.