Microexpression Detection Method Based On Deep Learning

Facial expression is an important expression of emotional information,and micro expression is a kind of facial expression that is not controlled by the nervous system.Usually,micro expression is expressed as personal real emotion,which makes micro expression detection widely used in public security and medicine.Because of the imperceptible characteristics of micro expression,it is very difficult to observe it by naked eyes,so some researches on micro expression focus on how to find and recognize micro expression.Most of the existing micro expression detection methods extract some motion features of micro expression,only considering the spatial characteristics of the features,but not considering the differences between the local features of space and time.Due to the complexity of the micro expression video,the spatial and temporal characteristics will affect the final micro expression detection effect.Generally,for a complete micro expression video,in terms of time dimension,only a few frames will have micro expression changes,and other frames may have changes caused by light,head movement,etc.In the spatial dimension,not all the region information in a picture is useful to us.On the other hand,this paper needs to do the following work to solve the problem of micro expression detection(1)A new feature learning method is proposed to transform the collected micro expression video into picture frames Sequence,the image frame is clipped to the specified size and aligned.The optical flow algorithm is used to calculate the optical flow image frame data set with the aligned image frame sequence.Each gray image can be regarded as a three-dimensional vector(x,y,t)through the optical flow algorithm,In a continuous change process,the motion change vectors of two frames are obtained,and the data set processed by the algorithm is transferred into a deep learning model,and more accurate motion features can be obtained by learning.(2)This paper proposes a feature extraction method(FLA-NET)based on time series information,which can make the extracted features have time sequence.It can learn features of continuous video frames in unsupervised learning mode.The proposed method can consider the spatial information of frame feature and add the feature change mechanism in time dimension.Add LSTM in Autoencoder model,According to the weight sharing characteristics of LSTM,the complexity of network model will not be greatly increased due to the increase of time sequence.The structure of LSTM time series model is added to make the network learn according to the time dimension.In this way,it is helpful to learn a deep neural network to calculate the motion feature information of each video segment,so as to complete the micro expression detection task.(3)A large number of experiments have been carried out on CASME?and CAS(ME)~2,which are two public datasets.The results show that the accuracy of micro expression detection is improved by adding temporal model.In the unsupervised training mode,the new model mechanism after the introduction of LSTM temporal network model can accurately extract the micro expression movement information in spatial and temporal dimensions.