Abstract
We propose an expression recognition method based on facial features from the psychological perspective. According to the American psychologist Paul Ekman’s work on action units, we divide a face into different facial feature regions for expression recognition via the movements of individual facial muscles during slight different instant changes in facial expression. This thesis starts from introducing Paul Ekman’s work, 6 basic emotions, and existing methods based on feature extraction or facial models. Our system have two main parts: preprocessing and recognition method. The difference in training and test environments, such as illumination, or face size and skin color of different subjects under testing, is usually the major influencing factor in recognition accuracy. It is therefore we propose a preprocessing step in our first part of the system: we first perform face detection and facial feature detection to locate facial features. We then perform a rotation calibration based on the horizontal line obtained by connecting both eyes. The complete face region can be extracted by using facial models. Lastly, the face region is calibrated for illumination and resized to same resolution for dimensionality of feature vector. After preprocessing, we can reduce the difference among images. Second part of our proposed system is the recognition method. Here we use Gabor filter banks with ROI capture to obtain the feature vector and principal component analysis (PCA) and linear discriminant analysis (LDA) for dimensionality reduction to reduce the computation time. Finally, a support vector machine (SVM) is adopted as our classifier. The experimental result shows that the proposed method can archive 86.1%, 96.9%, and 89.0% accuracy on three existing datasets JAFFE, TFEID, and CK+ respectively (based on leave-one-person-out evaluation). We also tested the performance on the 101SC dataset that were collected and prepared by ourselves. This dataset is relatively difficult in recognition but closer to the scenario in reality. The proposed method is able to achieve 62.1% accuracy on it. We also use this method to participate the 8th UTMVP (Utechzone Machine Vision Prize) competition, and we were ranked the second place out of 10 teams.