Research On Extraction Of Facial Expression Features And Hierarchical Classification

Facial expression recognition is an important branch of affective computing research field, and has great significance to the realization of human-computer interaction. The thesis analysis and research two aspects of facial expression recognition, one is feature extraction, the other is expression classification. Improve the traditional geometric feature extraction method and traditional texture feature extraction method. Present an effective feature fusion classification method to make full use of the self distribution of different features. The main work is as follows:(1)A facial geometric feature extraction method is proposed based on the similarity of neutral expression in view of the traditional geometric feature extraction algorithms are not robust to the differences of individual faces. In the process of expression recognition, first using multi-scale vector triangle patterns to measure the similarities of the test sample and the train neutral samples; second set an appropriate threshold to select neutral expression samples with high similarity, then using AAM (Active Appearance Model) to locate the key points in the face, calculate the proportion coefficient between the test sample and each similar neutral expressions respectively; finally calculate adaptive weighted geometric feature of the test sample according to the similarity, and use SVM (Support Vector Machine) to classify the features. The experimental results show that faical geometrical feature extraction method based on similarity of neutral expressions achieved better recognition performance than the traditional geometrical feature extraction methods, and reduce the influence of individual difference to facial expression recognition effectively.(2)To reduce the limitations of the traditional texture feature extraction method in expression recognition, a texture feature extraction method is proposed based on sufficient vector triangle. First, the eyebrows, eyes, and mouth in a facial expression image are identified and extracted to make the feature description more targeted; second, the local features from the organ images are then obtrained and processed by the expanded vector triangle pattern, in this way the outline and detail features of the images can be statistic; finally, different scales of sufficient vector triangle patterns are used to describe the features of the same organs, various scales of sufficient vector triangle patterns are then combined to describe the features of the same organ. The experiment results show that, texture feature extraction method based on sufficient vector triangle adequately expressed local texture features of images, and has lower time complexity.(3)According to the self distribution of different features, a hierarchical classification method is presented based on texture and geometrical features. First extract the global geometrical features and the local texture features respectively; second use different feautes to train different SVM classfiers; finally according to the self distribution of different features, determine expression categories by the local layer and the global layer, if the expression is unable to determine in the first two layers, it will accumulate to the mixed layer, in the layer the kind of expression is determined according to the maximum posteriori probability. The experimental results show that the hierarchical classification method effectively uses the distribution characteristics of the features itself, and the judgement has higher reliability.