Abstract: Facial expression is a natural, powerful and efficient mean of human communication. Different subjects and varying degrees of emotion lead to the spontaneous expression. Based on this difficulty, this paper established a probabilistic model between the facial action units (AUs) and the facial expression. In this model, the face was divided into two parts, eye brow area and mouth area, and the Gabor wavelet was used to perform the areas. Then AUs/expressions were recognized by a machine learning method which combined the K nearest neighbor (KNN) and the Bayesian network (BN). By training data and using priori knowledge to learn a model, this enhanced method provides AUs for different weights and select the optimal expression according to the maximum a posterior probability (MAP). Experiments illustrate that, the framework proposed in this paper showes a high recognition rate to different subjects and different degrees of emotion. It's an efficient, robust and automatic facial expression recognition system.