Research On Facial Micro-expression Recognition Technology Based On Temporal And Spatial Features

Micro-expression is a special facial expression with a very short duration and a low degree of facial movements.Compared with ordinary facial expressions,it can reflect the authenticity of human emotions and can be widely used in clinical medicine,judicial interrogation,and countries.Security and other fields.At present,micro-expression recognition mainly depends on manual identification,which has the problems of long training time,high cost and poor accuracy.With the development of artificial intelligence technology in recent years,micro-expression feature extraction and recognition has gradually become a research hotspot in the field of computer vision.A small number of researchers have made great progress in applying deep learning methods to micro-expression recognition,but due to insufficient micro-expression data,inconsistent image sequence length,and difficulty in extracting effective features,the accuracy of micro-expression recognition is still Low,it is difficult to meet the current industry application needs.To this end,based on the deficiencies of the existing micro-expression feature extraction methods,This thesis mainly uses the microexpression recognition algorithm of temporal and spatial features.The main research contents are as follows:The micro-expression data set is augmented by optical flow approximation and image reversal to solve the problem of overfitting caused by insufficient samples during model training.The local binary feature method is used to realize the global face feature point detection,and the spatial position of the face feature point is used to achieve the alignment and clipping of the micro-expression image.The frame insertion method based on convolutional neural network is used to interpolate the micro-expression image sequence,and the frame number is normalized through the TIM model,which effectively reduces the loss of micro-expression image features.Experimental comparison analysis shows that the method The generated intermediate frame is closer to the original frame,which is better than the traditional frame interpolation method.Aiming at the characteristics of micro-expression dynamic change in space and continuous association in time,the spatial feature extraction method based on Convolutional Neural Networks(CNN)and the Long Short Term Memory Network(LSTM))'S temporal feature extraction method to achieve accurate description of the spatial and temporal features of micro-expressions,a micro-expression recognition model based on temporal and spatial features is constructed.This model uses a convolutional neural network to extract the spatial features of each frame of micro-expression images in the video sequence.And input one by one to the long-term and short-term memory network to obtain the time-series dynamic change information between the micro-expression frames.Through the fully connected layer and the loss function calculation,the accurate classification and recognition of the facial microexpressions can be realized.Considering the effects of CNN types,the combination of CNN and LSTM networks,LSTM network parameters,image sequence length and other factors on the model recognition results,the network model with the best effect was obtained through three comparative experiments.The CASME II and SAMM data sets The recognition accuracy rate reached 68.19%.Finally,through comparison and analysis with four commonly used micro-expression recognition models such as LBP-TOP+SVM and DiSTLBP-RIP,the recognition accuracy and effectiveness of the model are verified.