Research On Cross-domain Expression Recognition Based On Domain Adaptation Method

Facial expression is one of the most important and essential way that the human beings to convey their emotional states.The study of facial expressions recognition(FER)is one of the most attractive research topics in affective computing and computer vision.with the efforts of many researchers,many meaningful research methods have been proposed in the research field of facial expression recognition.These achievements not only promote the development of psychology,neuroscience and early childhood education et al.,but also are widely used in many fields such as artificial intelligence,medical diagnosis,fatigue driving,business analysis,intelligent classroom,human-computer interaction and criminal lie detection in real life.In this paper,we focus on two challenging issues in facial expression recognition: cross domain facial expression recognition and cross domain facial micro-expression recognition.Based on the linear regression model and the dictionary learning model,several effective cross domain expression modeling and recognition methods are proposed respectively.On this basis,the proposed linear recognition method is extended nonlinearly,and a number of expression databases is adopted for experimental verification.The paper contains the following four works: 1.A transductive transfer learning linear regression model based on sparse(STTLR)isproposed to deal with cross-domain facial expression recognition tasks.The basic idea ofthis method is to eliminate the distribution difference between source domain and targetdomain samples by means of the transductive transfer learning model.To better eliminatethe distribution difference,this method takes part in the learning of regression parametermatrix from the target domain with some unlabeled samples as auxiliary sets and from thesource domain with labeled samples as new training sets.The trained regression parametermatrix can better integrate the feature information of the source domain sample and thetarget domain sample through the auxiliary set.The model has more powerfuldiscrimination,so as to better complete the domain adaptive expression recognition task.Inaddition,considering the structural properties of handcraft features,a group sparse strategyis adopted to obtain effective regional information for regression parameter matrix whileremoving redundant regional information and reducing over-fitting.2.A cross-domain expression recognition method based on transductive deep transfer learningnetwork(TDTLN)is proposed.The TDTLN network model uses the strong nonlinearrepresentation of deep learning and the idea of auxiliary from STTLR model to combinethe source domain data and part of the target domain data as training samples to learn theoptimal nonlinear discriminant features,so as to improve the label prediction value of thetarget domain samples.Among them,the sample label of the auxiliary set is optimized as apart of the network parameters,so that the cross entropy function of the source domain andthe regression function of the target domain can be calculated at the same time,so as toreduce the distribution differences between the samples of different databases as much aspossible.This is helpful for network parameters to learn more discriminative high-levelsemantic information,so as to better complete the task of classification.3.A domain adaptation method based on dictionary learning based method(UDADL)isproposed to deal with cross domain expression recognition problem.As a linear method,the model trains an effective discriminant dictionary from the perspective of capturingcommon feature information in the source domain and the target domain.This model alsocab be used to represent the data features in the source domain and the target domain.Inorder to make the marginal distribution of two new feature representations more close toeach other,UDADL method also introduces two linear projection matrices to project thecoding parameters obtained from dictionary optimization respectively,so as to make thecoding parameters of source domain and target domain have closer marginal distribution.In addition,because some parameters of the model do not have closed solutions,UDADLmodel further introduces shared dictionary and analysis dictionary to relax the relatedvariables,so that the model is solvable.4.A domain adaptive learning method based on kernel dictionary Learning(KDLDA)isproposed.This to investigate the performance of the non-linear domain adaptive frameworkin facial microexpression recognition tasks.UDADL method is a linear method based onsubspace learning.The KDLDA method extends UDADL method from linear to non-lineardomain by using kernel mapping function,that is,the expression samples in the source and target domain are mapped from low-dimensional input space to a higher-dimensional feature space through the non-linear transformation of kernel mapping,and then the optimal linear interface is obtained in this new feature space.In this high-dimensional space,the inseparable features become linear separable.