| In recent years,with the vigorous development of artificial intelligence technology,computer vision technology is also gradually showing the trend of blossom everywhere.The future computer will have the ability which identify and understand emotions,and holds the same EQ as human beings.Facial expression recognition has become the research focus and hotspot in the field of computer vision,which has been extensively studied by many scholars.In this thesis,the cross-domain facial expression recognition is based on transfer learning and the arousal-valence emotion model.Besides,we focus on the two-layer fusion facial expression recognition method based on TPCA(Transfer Principal Component Analysis)and the expression method based on SA(Subspace Alignment).Firstly,in the cross-domain facial expression recognition problem,TPCA method is adopted to reduce the difference of data between domains by using the commonality between different domains.The algorithm projects original feature onto intermediate common subspace which is used to reduce the dimension of feature and reduce the difference of data between domains.Meanwhile,the psychological research and the existing research have proved that the arousal-valence dimension satisfies the positive correlation.To take advantage of the positive correlation and the advantage between different features,two-layer fusion facial expression recognition method is adopted.Experimental results show that the TPCA transfer learning method can improve the cross-domain facial expression recognition effect compared with the traditional principal component analysis.Moreover,the results of using dimension correlation recognition will be further improved.Secondly,subspace slignment method is adopted to solve the problem that TPCA algorithm takes a lot of time to calculate intermediate common subspace.The algorithm projects the source data onto source and the target data onto the target subspace directly.It does not need to calculate the common subspace,and consumes less time.The contrast experimental results indicate the subspace alignment algorithm is less time consuming than the TPCA algorithm.What’s more,the proposed method can obtain better recognition results than traditional PCA and traditional feature fusion.Finally,a facial expression recognition prototype system based on two transfer learning algorithms and continuous emotion dimension model is designed and developed.The system includes image display,model loading,and dimension prediction and other modules.The actual test shows that the system can use the two transfer learning algorithms to the effective recognition the facial expression in the continuous emotional dimension model. |
