Research On Emotion Recognition Based On Facial Expression

A person's emotion is usually reflected in his gestures,facial expressions and phonetic semantics,of which facial expression is the most powerful way to express human emotion,and non-contact facial expression recognition is very important for the realization of natural and harmonious human-computer interaction.Up to now,most researches have focused on how to improve the recognition accuracy as much as possible,while ignoring the time efficiency of the algorithm.The general output of facial expression recognition are discrete class labels,and most systems and methods do not take expression intensity into account.To apply facial expression recognition to the realistic scene better,this paper will carry out research based on facial expression video sequences.This paper applied a spatio-temporal descriptor HOG3 D for behavior recognition to dynamic facial expression recognition based on video sequences.Although HOG3 D descriptors extracted rich dynamic facial expression information,the feature dimension was also up to thousands of dimensions.In order to solve the problem that the feature dimension was too high,a new feature selection method PCA-Relief F was proposed to reduce dimension in this paper.The feature selection algorithm not only effectively reduced the feature dimension,but also selected the distinguishing features according to the distinguish ability of the features.The algorithm was efficient for facial expression recognition based on off-line video sequences.In order to realize real-time and fine recognition of the expression video sequences finally,this paper recognized single facial expression image in the video sequences.In the preprocessing stage,this paper used Viola-Jones detection algorithm to detect the initial frame,and introduced the target tracking algorithm based on spatio-temporal context for eye tracking,and realized the real-time preprocessing of video sequences.This paper selected the HOG descriptor to extract the feature of the single frame image according to the recognition effect and real-time performance finally.Considering the amount of expression information carried by different facial regions and their contribution to classification was different,a sub-block adaptive weighting algorithm was proposed for weighted fusion of facial features in different facial regions.The experimental results showed that the recognition accuracy of the adaptive weighted method was higher than that of the unweighted one.Finally the recognition rate reached 95.71%,and the time precision of the algorithm was 26.8 ms/f.On this basis,we proposed a SVM probability model for fine expression recognition,which could not only output the discrete class labels,but also output the membership degree of the corresponding classes.The variables in the SVM probability model that represented the distance between each sample and the hyperplane was just able to represent the change of the emotional intensity.It was not only able to analyze the emotional intensity of a single emotion,but also to distinguish the strength of the same expression.