Research On Facial Micro-expression Recognition Technology

Micro-expression is a kind of expression that is hard to be detected by human eyes,weak and lasts less than half a second.This kind of expression will appear when the external stimulation is very strong and people wants to repress his real emotion as much as possible.It can't be controlled by human and reflects one's real thought.The recognition of micro-expression has the prospect that it would be widely used in the field of interrogation and national security.For the untrained ordinary people,it is difficult to recognize micro-expression,so researchers try to use computer technology to complete this task.There are some difficulties in automatic recognition of micro-expression: it is difficult to collect micro-expression data,there are few samples in existing datasets,and the number of every class of samples is unbalanced,which is not conducive to the training of classifi ers;Because of the short duration of micro-expression,low intensity of movement and small range of action,the common expression recognition algorithm is no longer applicable,so it is necessary to design a special feature extraction method for micro-expression.To solve these problems,in this paper we study the methods of preprocessing,feature extraction,feature dimensionality reduction and expanding the volume of micro-expression data.Firstly,considering the differences of illumination,face shape and pose among the samples in the dataset,three data preprocessing methods are studied,including histogram specification,facial landmarks detection and face registration.By comparing the performance of two registration algorithm,LWM and Procrustes ana lysis,we conclude that Procrustes analysis is more suitable for micro expression data.Then,a recognition algorithm for data in video form is proposed.This algorithm uses HWP-TOP to extract the feature of micro-expression.The dimension reduction algorithm based on Grassmann manifold is used to reduce the dimensions of features and make them easier to learn.The performance of several dimensionality reduction algorithms is compared by experiments,and the rationality of Grassmann manifold dimension reduction is proved.In addition,a recognition algorithm for micro-expression apex frame is proposed.It takes the local features of micro-expression into account,and uses local non-negative matrix factorization to extract the features.To solve the problem that the generalization ability of classifier is weak due to the small amount of micro-expression data,we need to create some new micro-expression samples with the help of a large number of macro-expression samples.For that we improve a macro-to-micro algorithm and make that suitable for non-negative feature used in this paper.Experiments show that these new samples can improve the recognition accuracy significantly.Finally,a deep neural network combining the static and dynamic features of micro-expression is proposed.Firstly,CNN is used to extract the static features of the apex frame of micro-expression,and then GRU is used to process the feature time series composed of each frame of image features to obtain the dynamic features of expression.Finally,the two features are fused and recognized.Experiments show that the performance of recognition based on fusion feature is better than that based on single feature.