Cross-Domain Facial Expression Recognition Based On Transfer Non-Negative Matrix Factorization

Facial expression recognition(FER)usually refers to the use of recognition algorithms and human emotional knowledge to train computers to automatically recognize and judge the emotional expression of people in pictures or videos.In recent years,it has become one of the hot research topics due to its huge application potential in human-computer interaction,wisdom medical,online education,investigation,entertainment,and other fields.Traditional facial expression recognition algorithms usually assume that the data follow an ideal feature distribution.Therefore,most algorithms are experimented with and evaluated on a single database that follows the same feature distribution,to achieve satisfactory results to a certain extent.In real life,face images are often collected from different situations,such as ethnicity,collection environment,collection device,gender,age and so on.These different factors lead to different feature distributions of facial image features.Therefore,the performance of traditional facial expression recognition algorithms will reduce significantly influenced affected by these different factors.To handle the problem,this article proposes cross-domain/cross-database facial expression recognition algorithms based on transfer non-negative matrix factorization.The core ideas of this paper are as follows:(1)Firstly,from the perspective of transferable discrimination learning,we proposed a transfer discriminant non-negative matrix factorization(TDNMF)method.TDNMF uses a pair of similarity and dissimilarity information obtained from the label information to guide matrix factorization and feature transfer.At the same time,the method designs a graph Laplace regularization term,which constructs the nearest neighbor graph by inter-and intra-database similarity information,to preserve the local geometric relationship of features.Finally,this method will learn a common low-dimensional representation for the source domain and target domain.By comparing the proposed method with a series of contrast methods,it is proved that TDNMF can achieve better recognition results.(2)Secondly,based on TDNMF,we propose a transferable non-negative feature representation(TNFR)method to learn transferable features.The method first embeds a transfer reconstruction item into the graph non-negative matrix factorization framework.It uses the base matrix of the non-negative matrix factorization as the transfer matrix to reduce the distribution difference between the source domain and target domain.Secondly,the method designs a block-label regression term to increase the discriminant ability of the model.Different from traditional label regression,block-label regression can relax the strict dimensional constraints of traditional label regression,making the proposed model more flexible.Finally,the method is compared with several deep learning and traditional methods to verify the effectiveness.