Abstract
In this thesis, we propose a new facial expression recognition algorithm based on supervised locally linear embedding (SLLE) analysis on the optical flow and ratio image. In this algorithm, we first extract the face region from the face image to remove factors due to global head motion. Secondly, we compute the optical flow and ratio image between the neutral face and expression images and then apply the SLLE to extract the low-dimensional discriminating features from the expression motion and brightness variation. Thirdly, we compute the distance between the low-dimensional feature vectors to recognize the facial expression. Finally, we combine optical flow and ratio image properly to improve the facial expression classification. The experimental results on the JAFFE face database show the proposed algorithm outperforms the previous methods for facial expression recognition. We also use the Yale Face database for testing the expression recognition system trained from the JAFFE database. It still has good performance on the expression recognition rate. Therefore, we successfully use different database for testing and the result is comparable with testing on the same JAFFE database. The result shows that the system not only works well on JAFFE database but also has good performance on the Yale Face database.