Research On Expression Analysis Method Based On Ordinal Regression

Emotion plays an important role in the process of human information communication.Human emotional state is usually the comprehensive expression of body language,facial expression and voice semantics.As the most primitive way of communication,facial expression is the most natural and powerful way for human beings to convey their emotional state and intention.Facial expression analysis is the basis of emotion calculation and the key technology of human-computer natural emotion interaction.Facial expression analysis is mainly composed of facial expression recognition and expression intensity estimation.The work of expression recognition is mainly to recognize six kinds of basic expressions,including anger,disgust,fear,happiness,sadness and surprise.Expression intensity estimation further distinguishes the subtlety of intensity between similar expressions.Most of the existing work focuses on facial expression recognition,and there is little research on the estimation of facial expression intensity.However,just classifying basic expressions does not fully understand people's emotions.In order to further understand people's emotional state and intensity,expression intensity estimation has attracted extensive attention.Although researchers have invested a lot of energy in the field of expression analysis and created a lot of meaningful research results,there are still some difficulties in this field due to the subtlety and complexity of facial expression.It is difficult to train a model with high robustness and strong generalization ability and prone to over fitting problems.Especially in the aspect of expression intensity,there is still a lack of a large number of labeled data,so it is difficult to train the model with a supervised method.Although the ranking method can solve this problem,it can only estimate the relative intensity of expression,but not the absolute intensity of expression.In view of the above problems,this paper mainly studies the expression analysis algorithm based on order information for joint expression recognition and intensity estimation,and further proposes an expression intensity estimation algorithm based on convolutional neural network which integrates sorting and regression.The main work of this paper can be summarized as follows:(1)A sort convolution neural network(Rank-CNN)is proposed to analyze facial expressions using the sequence information of facial expression sequences.It unifies facial expression recognition and intensity estimation into a Rank-CNN framework.It increases the differences between different expression categories through intensity sorting,and reduces the intra category differences of the same expression through depth difference features,so as to be interested in the face The facial expressions were classified.In addition,any intensity of facial expression can be classified by comparing with the intensity of neutral expression.(2)A semi supervised framework of convolutional neural network(JRR-CNN)based on fusion sorting and regression is proposed to solve the fine-grained expression analysis problem.The over fitting problem is solved by constructing a lightweight network.At the same time,the absolute intensity estimation of expression frame is improved by introducing regression branch.In addition,it achieves the best effect on many public facial expression databases(PCC,ICC and MAE are 0.6551,0.5293 and 0.9241 respectively on the panel data set,and PCC,ICC and MAE are 0.7391,0.7216 and 0.1875 respectively on the CK+data set).The experimental results show that the method proposed in this paper effectively solves the problem of insufficient strength mark data and lack of good feature representation,and improves the performance.The results of this study will help to better understand human emotional state and promote the application of facial expression analysis in related fields.