Abstract: It is well known that the two-dimensional facial expression data contains limited information, and the poor performance of the facial expression recognition under the condition of changing illumination and posture. In order to overcome these shortcomings of the 2D facial expression, In this paper, we propose and explore a novel method to recognize human facial expression in 3D based on Self Organizing Map(SOM). In the method, the mean and variance are used to describe the convex and concave surface of the face which become the facial expression change characteristics datas. The simulation experimental results showed that the effect of using the classification and recognition of SOM network was superior to the AdaBoost algorithm.