Vision Research And Implementation Based On Facial Expression Recognition

Emotion is an important part of human intelligence. Accurate identification of human emotions helps the computer grasp the human state of mind and serve humanity better. This paper carried a study on emotion model from the facial expression recognition. The detailed contents and key innovations of this thesis are listed as follows:1. Acquiring and preprocessing expression image. There are two sources of expression images:collected by network cameras in real-time and facial expression database. In this paper, we captured facial expression image in real-time using webcam, and have data pre-preprocessed about collected images and database images. Expression image preprocessing includes face detection and localization, posture correction, image filtering and image normalization.2. Expressions are recognised by texture features and geometric features respectively. About texture recognition, we map eigenvector into the nonlinear high dimension space using the KPCA and process eigenvector by principal component analysis. The method achieves the purpose of dimensionality reduction while enhance the distinguishability of different categories. About geometric feature recognition FCBR, which combines fuzzy logic and case-based reasoning, was taken to blur face geometric features, and then establish case base, we established case-retrieval system according to "if-then" rules. This method can improve accuracy of expression recognition through simulating human thinking. Experimental results based on JAFFE database show that those two improved or fusion methods can improve the rate of recognition.3. Using fuzzy integral method, we fused the recognition results from texture and geometric features in the decision level. This method takes recognition rate of texture and geometric method into account and gives the final recognition result from the view of probability. Experimental results based on JAFFE database show that the fusion method can give a higher recognition rate.4. Using the K-means clustering method to measure the intensity of expression. The paper introduces the expression intensity level to measure sequences expression intensity, and train the geometric eigenvector of the image through cluster analysis, then establish expression intensity model and evaluation criteria. The results obtained by expression sequence show that the proposed method has better strength recognition.5. Establishing emotional model based on changes of emotional state. The model incorporates the following factors:external stimuli promoting the role of emotion update, interactions between emotion, damping of emotional components, and the influence of personality on emotion update. Experimental results show that the emotion model can qualitatively and quantitatively determine external emotional stimuli according to the beginning and end emotional state and the individual personality factors.