defines one primary direction. Second, the expression recognition system uses the 15 feature vectors for facial ex-This paper introduces an automatic facial expression recognition system which consists of two parts: facial feature extraction pression categorization. He showed an accuracy rate of and facial expression recognition. The system applies the point nearly 80% for recognizing four types of expressions: hapdistribution model and the gray-level model to find the facial piness, anger, disgust, and surprise. Morishima [8] develfeatures. Then the position variations of certain designated oped a facial emotional model which employs a representapoints on the facial feature are described by 10 action parametion facial feature action based on the description of the ters (APs). There are two phases in the recognition process: epic of facial expression. Facial expression is described as the training phase and the recognition phase. In the training the point of the space and the face animation can be dephase, given 90 different expressions, the system classifies the scribed by the locus in 3D emotion space. principal components of the APs of all training expressions Yacoob and Davis [3] proposed an approach for analyzinto six different clusters. In the recognition phase, given a ing and representing the dynamics of facial expression. facial image sequence, it identifies the facial expressions by extracting the 10 APs, analyzes the principal components, Their system consists of locating of tracking prominent and finally calculates the AP profile correlation for a higher facial features, optical flow analysis, and the classification. recognition rate. In the experiments, our system has demon-It achieves a recognition rate above 80% for all six expresstrated that it can recognize the facial expression effecsions. Rosenblum et al. [4] extended the work of [3] by tively. Β© 1997 Academic Press using a connectionist architecture. Individual emotion networks were trained by viewing a set of sequences of one emotion for many objects. The trained neural network was