Research On Intelligent Recognition Method Of Static Facial Expression

In recent years,with the development of artificial intelligence technology,affective computing has attracted more and more attention.As the core technology of emotional computing,facial expression recognition has high research value and engineering practical significance.Facial expression recognition can be divided into static expression recognition and dynamic(sequence)expression recognition.Static expression recognition refers to the judgment and recognition of facial expressions based on static facial images.Due to the low complexity of static expression recognition model,it has widely landed in security monitoring,autopilot,business strategy and other industries,and has become a hot topic of academic research.Traditional static facial expression recognition methods rely on image processing algorithms,which have low recognition accuracy and weak generalization ability.At present,the intelligent expression recognition method based on convolution neural network has gradually replaced the traditional expression recognition method.However,the accuracy of intelligent expression recognition method is limited by the quality of data set and network model.The existing facial expression recognition data sets have some shortcomings,such as small number of samples,poor quality of samples,unbalanced samples,inaccurate labels and so on,which seriously affect the accuracy and generalization ability of the training model.Therefore,the practical value of the model trained based on the above data sets is low.In order to solve the problem of data set quality in static facial expression recognition,this thesis constructs a high-quality facial expression recognition data set AffectRAF considering sample balance,taking into account both sample quality and sample number.At the same time,based on the constructed data set,a static expression recognition network SRINet based on deep learning is proposed,which improves the accuracy of recognition.The main innovations and work of this thesis are summarized as follows:(1)A sample balanced data set AffectRAF is constructed.The existing open facial expression data sets have some problems,such as unbalanced samples,limited sample acquisition conditions,uneven sample quality,and limited generalization ability of training effect.Therefore,this thesis constructs a data set AffectRAF,which takes into account sample quantity,sample quality and sample balance,and optimizes the data set.(2)A deep neural network model SRINet is proposed.According to the characteristics of AffectRAF data set,a static facial expression recognition model SRINet is proposed,which is suitable for AffectRAF.The lightweight structure of multi-layer small convolution kernel is used to effectively extract facial expression features and improve the accuracy of static facial expression recognition.(3)This thesis verifies the effectiveness of AffectRAF and SRINet.The experimental results show that the recognition rate of the classical facial expression recognition model decreases greatly on AffectRAF,which shows the necessity of constructing AffectRAF data set.At the same time,the proposed SRINet network shows superior performance on classic expression recognition data sets such as Fer2013,JAFFE,CK+and AffectRAF,which shows that SRINet has strong model generalization ability.