The invention relates to in an analysis method of the Facial Expression image classification and semantic evaluation on the basis of semantic evaluation and the quantization. This invention makes use of the sample collection of training Facial Expression images, extracting expression characteristics to form a marker map of LG Vector, projecting it to the Principal Component PCA subspace, making the use of these reduced-order LG Vectors to learn mixed multi-dimensional t-distribution, as the semantic judgment of six basic emotions of the images, in according to that which probability taken up by expression is maximal, that image is judged thereby to that expression. Which resolve the recognition difficulties of the uncertainties, automatic facial expression existed in the current technology, and the singularity difficult to overcome such defect. This invention is an extremely flexible and powerful modeling tools based on the statistics, provides a more robust ways, avoiding the extreme estimates of posterior probability of the subordinate component of sample observed, the training samples do not require labeling, do not require any post-processing, less sensitive to outliers, avoid this artificial Judgment.