Unsupervised Cross-domain Facial Expression Recognition Based On Transfer Learning

Facial expression recognition is a research hotspot in the field of emotion computing and computer vision.In recent years,with the rapid development of artificial intelligence technology,expression recognition,as an essential part of intelligent human-computer interaction,has a wide range of application prospects in the fields of multimedia entertainment,human-computer interaction,machine intelligence,etc.,which has attracted the extensive attention of researchers and put forward many effective methods.However,most of the existing facial expression recognition needs to meet the requirement that the data set used for training and testing is the same data set,that is,the expression images in the training set and the test set are collected under the same conditions,so they are subject to independent and identical distribution.However,in many practical applications,this assumption does not hold,because test data sets are usually collected online and more difficult to control than training data.For example,through different devices,from different angles,lighting conditions and background,the collected data samples usually have different feature distribution,so the traditional expression recognition method is no longer applicable,which will lead to these algorithms can't achieve better recognition results in cross domain expression recognition.Because,at this time,the optimal error estimation on the training data is no longer the optimal estimation of the test data error.In this paper,we mainly study the problem of facial expression recognition in which the training set and the test set come from two different databases,and there is no label data in the test set.We define this problem as unsupervised cross domain facial expression recognition.The databases used for training and testing are called source domain and target domain respectively.Therefore,in order to solve the problem that the training data and test data are subject to different distribution in the practical application of expression recognition.In this paper,we first propose a transfer learning algorithm based on distributed alignment(Da L)for crossdomain expression recognition.In this method,the source domain and target domain data are mapped into a common subspace by learning a feature transformation.In this subspace,the maximum mean difference MMD without parameters is introduced to measure the distribution difference between different domains,and the marginal distribution and conditional distribution of the same source domain and target domain are combined to reduce the distribution difference between domains.Secondly,a subspace alignment algorithm(SA)based transfer learning is proposed for cross-domain expression recognition.The algorithm maps the source domain and the target domain to their respective subspaces,and then by learning a linear transformation matrix,the source domain and the target domain subspaces are directly aligned to reduce the distribution difference between the domains.The experimental results show that the two algorithms are effective in cross domain facial expression recognition.Then,in view of the limitations of the distributed alignment and subspace alignment algorithms,a joint subspace and distributed alignment algorithm(JSDA)is proposed,and the algorithm is compared with Da L and SA in different experimental settings.The experimental results show that the recognition effect of JSDA on different expression databases is better than Da L and SA.Finally,we summarize the research content of this article and look forward to further research in the future.