Research On Video Facial Micro-Expression Recognition Based On Deep Learning

Face micro-facial recognition has become a hot topic in the field of industrial intelligence and pattern recognition.Facial expression is the true expression of people's inner emotions,research facial micro-expression can make the machine a better understanding of human emotional information in human-computer interaction to better communicate with people to help people work.In this paper,we study the recognition of more subtle expressions on the basis of expression recognition,and divide the microexpressions into seven categories: happy,sad,surprised,aversion,fear,pain and neutrality.Micro-expression changes generally do not exceed 0.25 s,the study of micro-expression video stream more than a single image to express the emotions of people.Therefore,this paper proposes some new methods for video face micro-facial expression combining with depth learning and experiments.The main work of this paper:1.Establish micro-expression database.First,a set of sample datasets suitable for the proposed network model is established,which is divided into a color and grayscale micro-facial expression dataset,and 12 sequences are taken as the sequence length of one expression in the microexpression video sequence.2.In this paper,we propose a new method to identify the micro-facial expression of the video using the convolutional neural network with depth learning.Three microexpressions are sent to the convolutional neural network at the same time,one step at a time,so the constructed network can guarantee the simultaneous micro-facial expression sequences.In this paper,different convolution network models have achieved good experimental results.3.Simultaneous input of convolutional network Although the sequence of samples is inputted at the same time,all the sequences are identified,but the characteristics of time are neglected and the computation is large.In this paper,recursive neural network is used to deal with time series features effectively.A network model of CNN-LSTM is proposed by combining it with convolutional neural network.Experiments show that the network model can identify the time-series at each time point,and Have very good classification performance.In this paper,two experiments are carried out on Caffe and TensorFlow,respectively.These two frameworks can promote the deep learning and have the visualization tools.Experiments show that the two depth learning models of this paper have advantages over other algorithms,better effect and higher practicability.