Recognizing faces with expressions: within-class space and between-class space

We propose a technique for expression invariant face recognition, which is different from the eigenfaces method from two aspects: the first is that instead of applying principal component analysis (PCA) on the pixel domain to obtain eigenfaces, we train eigenmotion by applying PCA on motion vectors obtained from the training face images with expression variations; the second is to consider the reconstructed errors of a test image in two spaces: the between-class eigenmotion subspace and the within-class eigenmotion subspace, which are used as the classification rule, in contrast to the traditional methods such as Euclidean distance or Mahalanobis distance in one subspace. Experimental results show that this method performs better than the eigenfaces method in the presence of facial expression variations.