Expression Recognition Algorithm Based On Face Sub-Region Weighting And LDA

Facial expression is one of the most important manifestations of people's daily communication which has great significance to the research and development of emotional computing.For static expression images,the matching method of single whole template contains more unrelated regional feature.The feature dimension is high which making it difficult to achieve the better recognition results.This thesis starts from the view of geometric confidence interval and do the research combined with linear discriminant analysis algorithm.The main work is as follows:(1)A weighted fusion feature extraction algorithm based on subarea(confidence regions)and Multi-feature is proposed.Due to the areas that are not related to the face part for the detection area,we can get a more accurate face and its sub-area by giving a clipping strategy based on geometric a priori to the detection area.For the facial images existed expression-independent area and single feature depicts inaccurate,the thesis extracts the feature of the face region and confidence interval by using Gabor wavelet and HOG.Then experiments are verified by studying the a priori information(sensitivity)of the confidence interval in the facial expression.Finally,the corresponding weights for different confidence interval are set,and the weighted fusion features are gottern.According to the experiments on different data sets,this algorithm is proved effectively.(2)An improved linear discriminant analysis algorithm based on within-class scatter matrix correction is proposed.To overcome the lack of discriminanting characteristics for the initial dimensionality reduction feature,by introducing information on cosine similarity in the class divergence matrix,the method of using angle cosine between each sample vector and its mean vector multiplied by the linear transformation to the corresponding covariance matrix has achieved a better intra-class polymerization and inter-class dispersion.According to the experiments on different data sets,this algorithm is proved effectively.(3)A GRNN neural network classifier applying to the field of facial expression recognition is constructing firstly.For the limitations of traditional classifier for nonlinear sample fitting of small samples,this thesis construct a GRNN classifier and embedded in the expression recognition algorithm through the analysis of the facial expression data.The algorithm uses fusion feature as the network's input.The training is completed after the training data passed through model layer and the summation layer.According to the experiments on different data sets,this algorithm is proved effectively.