Micro-expression Recognition And Visualization Based On Convolutional Neural Network

Facial expression,the true notions of human which revealed through the changing muscle in the eyes,lips and other parts of the body,is the direct medium to communicate and transmit all kinds of information of people.Unlike conventional facial expressions that people know,micro-expression is a unique way of expressing emotions,it cannot disguise and hide,and can be expressed of the true intentions of people in the most direct and clear way,therefore,it has great research value in many areas like public safety,medical clinic and social life.There are three distinct features of facial micro-expressions:"Small action range"?"Short duration" and "Cannot be present in multiple facial areas at the same time".In order to solve the three problems above and improve the practical application value of micro-expression recognition,we mainly completed the following three aspects in this thesis:First,to solve the problem of "small range and difficult to be recognized",in this thesis,we proposed a video motion amplification method based on the elimination of eye interference,which solved the problem of video distortion caused by blinking during amplification,in the meantime,the existing data are amplified one by one and reconstructed as a new enlarged data set named as mag CASMEs mag_CASME?.Secondly,the micro-expression recognition technology automatically implemented by the computer is still limited to the difficulty of improving recognition accuracy,which is why a face recognition method based on convolutional neural network is proposed in this thesis,and is successfully realize the micro-expression recognition using neural networks.In the meantime,this thesis eliminates the traditional method which combining data labels to improve the accuracy and realizes the full classification of micro-expression.Thirdly,according to the characteristics of,"local appearance" of micro-expressions,we introduce the concept of CNN to realize a visualization method of movement of facial micro-expression based on optical flow estimation.An optimized optical flow estimation effect was obtained through fine processing of local optical flow characteristics.At the same time,the corresponding color strategy was used to assign specific colors to optical flow quantities in different directions and amplifications,so as to realize the visualization of micro-expression actions.