Research On Some Key Technologies Of Facial Micro-expression Recognition

As a barometer of human emotions,facial expressions reflect the joy and sorrow of human beings.Over the years,humains have devoted themselves to expression recognition to study their emotional changes,especially for face recognition,expression recognition and other related tasks intelligently.As a real and unconscious expression,spontaneous expression can reflect the true emotions of human beings accurately and is favored by experts and scholars gradually.As a kind of spontaneous expression,micro-expression can reflect the real emotional changes of human beings.It is a highly effective and accurate behavioral clue,so it has widespread application value in lie detection,counter-terrorism,trial,clinical diagnosis and so on.However,since the micro-expression is low in intensity,short in existence,and may only appear in part of the face area,it is difficult to recognize the micro-expression automatically.In order to improve the effectiveness of micro-expression recognition,this paper has carried out studies from the following three aspects:1)Expounding the current research status of micro-expression recognition,including the basic flow of micro-expression recognition,the development process of micro-expression recognition method and the introduction and comparison of micro-expression databases.2)A micro-expression recognition method based on deep temporal-spatial feature fusion is proposed.In the algorithm,a global feature descriptor for describing the facial micro-expression is proposed.The spatial and temporal features are extracted by the spatial convolution neural network and the temporal convolution neural network respectively,and the two features are merged through the spatio-temporal feature fusion layer.The resulting global features are used for micro-expression prediction classification.3)A micro-expression recognition method based on residual network and Long Short-Term Memory is proposed.Firstly,in order to better describe the characteristics of micro-expression,we use optical flow images as input data.Then,the residual network is used to extract the spatial features of the input data,and then the extracted spatial features are sent to the Long Short-Term Memory(LSTIM)Network for temporal learning.Finally,the softmax classifier is used to classify the extracted feature vectors.