Facial Expression Feature Modeling Algorithm And Application

Human facial expressions contain rich emotional information.By observing the changes in facial expressions,a person’s emotional state can be judged.With the development of modern computer technology,people have put forward higher and higher requirements for the visual design of human-computer interaction interfaces,making facial feature modeling and expression generation technology play a very important role in human-computer interaction.How to design a realistic facial expression generation system is an important part of studying facial expression feature modeling and expression generation.Based on the analysis of the change law of facial expression features,this dissertation studies the regression-based facial expression feature modeling algorithm to generate realistic facial expressions.The specific research content is as follows:(1)Research on the related technology of facial expression generation,and discuss the moving least squares(Moving Least Squares,MLS)and moving regularized least squares(Moving Regularized Least Squares,MRLS)based on feature point control.A regression equation is established for the feature points of the face,and the control of the vertices of the mesh is used to achieve the effect of image transformation.Since the MRLS algorithm adds a regular term on the basis of the MLS algorithm,the least squares method based on moving regularization can realize the non-rigid transformation effect of the image,which is better than the MLS algorithm.(2)Expression modeling.Based on the changes between any of the six basic expression(anger,disgust,fear,surprise,happiness and sadness)features and the neutral expression features,the associate models are established and used for expression prediction.First,a linear regression model is established according to the coordinate changes of the six types of facial expressions and neutral expression features.Since the linear regression model cannot effectively predict the expression features,it is proposed to use the time-varying nonlinear KNR(Kernel-based Nonlinear Regression,KNR)model to predict the expression features,and use the MRLS algorithm to generate expressions,and calculate the absolute error and mean square of the two models at the same time Root error.The experimental results show that the KNR model is better than the linear regression model in predicting facial expression features,and the generated facial expressions are also close to the actual facial expressions.(3)According to the change process of the six basic expression sequences in the time series,six kinds of expression models are trained,so that when the first two frames of expression features of any person are given,the expression of the next frame is predicted and generated.First,the features of the expression sequence are obtained by the coordinate difference of the feature points between the facial expression image and the neutral image in the expression sequence,and six auto-regressive models(Auto-Regressive,AR)and KNR models of the expression sequence are obtained by training.Given the features of expression sequences,the model obtained through training can generate six expression prediction sequences.Since the AR model has a worse and worse prediction effect with the change of expression amplitude,it is further proposed to use the KNR model to learn the displacement changes of expressions in time series.By comparing the experimental results of the two models,it is found that the KNR model is not affected by the magnitude of expression changes,and the prediction effect is better than the AR model.In summary,the KNR model can well express the process of facial expression feature change,and has important value for predicting and generating facial expressions.