The Method And System Research On Automatic Facial Expression Calssification

In daily life and social intercourse, facial expressions can dynamically reflect one's emotion state and perceive state. Therefore, its analysis has received considerable attention in the fields such as business negotiations, children's characters analysis, criminal inquest, and patients' psychological state estimation. In addition, it is also very important in the development of visual-phone and human computer interaction system (HCIS).With the development of image analysis, computer vision, and pattern recognition, the computer science based automatic facial expression classification has gained extensively progress in resent years. This dissertation focused on dealing with the key problems in accomplishing automatic facial expression classification. The main research contents are listed as follows:1. A novel face detection method,which combined the merit of neural network and wavelet,was proposed. Two stages were involved in this method. In the training phase, some background masks were generated and added to the background of the AR face database examples. Then the pretreated AR face images were decomposed by the wavelet transformation and the corresponding wavelet coefficients were set as the input samples for a well-designed neural network. In the face detection phase, the neural network was used to judge whether the input region of the resized images contained a face. The detection results were validated and combined by the rule-based method and SUSAN (Smallest Univalue Segment Assimilating Nucleus)-based method. The experimental results showed that this method could accomplish detection well and be robust to both lightness variation and noise.2. A facial feature points matching and prediction method, which was based on the combination of pyramid decomposition and wavelet moment, was presented. The wavelet moment method performed the multi-scale representation of image information while pyramid decomposition method achieved the connection between the local information and global information. Finally, the reliable facial feature points matching and prediction results could be got by the proposed method.3. An efficient method, which could characterize the facial deformation using the eigen-modes, was designed. While the medial parts of the eigen-modes were selected, reliable deformation information of the facial regions could be extracted. After the regularization method was used, the coefficients in the new body-centered coordinate system could be calculated. The modulus of these coefficients composed the featurevector and they were used for automatical facial expressions classification.4. A method, which can efficiently compute the two-dimensional Tchebichef moments, was proposed. First, a fast one-dimensional Tchebichef moment calculating method was deduced by using the Tchebichef recurrence formula. Then, it was extended to the two-dimensional Tchebichef moment's computation. Compared with the direct method, the experimental results showed that this new algorithm could greatly reduce the computational complexity .5. A neural network tree (NNTree) based facial expression classification approach was developed. It introduced an NNTree method to implement feature selection and facial expression classification. The overall structure of an NNTree is a decision tree (DT) whose internal nodes are embedded with modular expert neural network (ENN). In facial expressions classification, the experimental results showed that this method was more efficient than the multi-level perceptron or decision tree. The practicability of the neural network tree, the influence of the neural network tress's complexity and its ability in efficiency feature discovery were also discussed.