Face Expression Recognition Based On Singular Value Decomposition

Face Expression Recognition (FER) system is expected to have numerous applications in many fields. The whole system can be divided into several parts: expression image capture, pre-processing, face feature detecting and locating, face segmentation and normalization, expression feature abstraction, and face expression recognition.In this paper, the face images used are JAFFE from Kyushu University of Japan. These images are highly standardized, and we know them well. So, after cutting, scaling and normalizing these images by hands, we directly do the job of extracting the feature vector and recognizing face expression. First, we give a short review on the recent development in FER research, and then introduce the kinds of popular methods of FER of the typical algebraic feature (Singular Value feature). Based on Singular Value Decomposition (SVD), we start the FER research. Compared to the traditional SVD method, the face expression image is projected onto two compressed orthogonal matrixes, which come from the SVD of the standard face expression image obtained by averaging all training samples and then the projecting coefficients are used as the algebraic feature of the face expression image. The robustness of this feature is proved and used for FER. In matching stage, the traditional Nearest Neighbor Classifier (NNC) is improved to recognize the unknown faces expression by using Euclidean distance as the similarity measurement. Another method is also introduced, because the Discrete Cosine Transform (DCT) is good at energy concentration, data de-correlation, and rapid calculation, it can be a good method of feature abstraction. So a kind of FER method based on DCT and SVD is tried in this paper. After dividing the face image into 8×8 sub-blocks, do DCT to every block, and extract the DCT coefficients by two methods: within a square area, or by a zigzag scan. All the DCT coefficients are organized into a whole coefficient image by a kind of combination before SVD, and the improved NNC is used to recognize the unknown face expressions with the Euclidean distance as the similarity measurement. Experiment results show that both the recognition ratio and the recognition speed of this method, compared to traditional SVD, are improved much. But the method of using averaging face expression feature may be short of generalization, which still needs improvement.