The article proposed a new way of facial expression recognition by reconstructing the PCAmethods. Separating the training set into 7 parts by expression, and get the Orthonormal Basis of each subsetexpression by PCA algorithm. Then make projection with the original image on each of the OrthonormalBasis, rebuilt the projection coordinate and recognize the expression by the combination of D-valuesummation and similarity methods. Finally, we get the final result by comparing the similarity between theoriginal and reconstruction image.