Study And Implementation Of Facial Expression Recognition System For Sentiment Analysis

Facial expressions have rich information of human behavior.It is the most important carrier of emotion.As an important branch in emotional computing,expression recognition is not only the foundation of sentiment analysis,but also the precondition for computer to understand human emotion.In the past,facial expressions recognition research work just considered variation based on Geometric features,texture features,custom templates,and some other pattern recognition technologies.These methods only consider parts of the characteristics of facial expressions.Due to the key point of this way is not to extract the comprehensive and efficient feature of facial expression,it missed some details of the facial expression.Moreover,it performs just not robust about light changing,people changing,race changing,etc.Based on above facts,a new feature extraction method is proposed in this thesis.And it reaches a better recognition rate combined with multi-class classification method.A new facial expression feature extraction method based on combining HOG descriptors and the Autoencoder of deep belief network is proposed in this thesis.Firstly,HOG feature of the image is extracted through the image unit division,and the combination of different scale operators and overlapping method.Then,deep belief networks abstract different levels of expression representation by the mid-level,and add sparse restrictions on it,which regard as a feature to describe the image.After that,the new feature is extracted by combining two different representations of face expression characteristics.And the feature extraction method verified by experiment can not only reflect the shape variation of human face,also can describe the details of expression.It is worth mentioning that the new feature has certain generalization ability and excellent characteristic express ability.In addition,by leveraging the feature extraction method in this thesis,the multi-class classification model of SVM is utilized to recognize the expression of human face,which considers the relationship between the expressions categories,it can summarize the rules of each category according to the experiment result.The comparison between proposed method in this thesis with other classification methods based on the JAFFE and CK+ facial expression image datasets shows that the method performs better than some baseline methods,which validates the effectiveness of proposed method in this thesis.In the end,by integrating the expression image preprocessing module,feature extraction and face recognition methods proposed in this thesis,a facial expression recognition prototype system is implemented under the Matlab simulation platform and the Eclipse integrated development environment.