Virtual Environment Interaction System Based On Micro-expression Recognition

Facial expressions contain a large amount of emotional information and are the main research topic of emotion recognition.Accurately identifying facial expression categories can effectively promote people's daily communication.As a special facial expression,micro-expressions are difficult to be captured and accurately recognized by people.However,it is difficult to hide and forge micro-expressions,which has broad application prospects in the fields of security lie detection,business negotiation,and educational evaluation.Therefore,automatic recognition of micro-expressions is of great significance.In the early days,a large number of researches on micro-expression recognition were based on traditional machine learning,and micro-expression features were extracted through artificially set manual features.This method can only extract shallow features of micro-expressions,and cannot fully extract the space and time domain of micro-expressions feature.In recent years,convolutional neural networks have performed well in two-dimensional image recognition tasks,and the recognition rate has surpassed traditional machine learning methods.Therefore,a large number of researchers have used convolutional neural networks to study micro-expressions and obtained a better accuracy rate.This paper studies micro-expression recognition based on deep learning methods.Through different pre-processing methods,a 3D-Dense Net model is constructed and the attention mechanism is fused to extract features from micro-expressions,therefore got better recognition results.Moreover,use this micro-expression recognition algorithm model to build a real-time recognition system of micro-expression based on a virtual learning environment.The main work of this article are:1.In terms of pre-processing of micro-expression data,this paper takes a variety of pre-processing measures.In response to the problem of the small sample size of micro-expression datasets,this article mirrors and rotates SMIC,CAS(ME)2,and CASME-II micro-expression datasets and expands them to 10 times.Using 3D-FFT to extract micro-expression sequences apex frame,and normalize the micro-expression sequence to 20 frames,including the apex frame.Eulerian video magnification processing is performed on the normalized micro-expression image sequence to increase the movement amplitude of the micro-expression.2.Constructed a 3D-Dense Net model for feature learning and classification of micro-expressions.For SMIC,CAS(ME)2 and CASME-II datasets,in this paper,the best recognition accuracy rates are 95.12%,92.63%,and 82.74%,respectively,and using 10-fold cross-validation test the validity of the model.Besides,the attention mechanism is integrated.By squeeze and exciting the 3D-Dense Net network channel,the recognition accuracy in the SMIC,CAS(ME)2,and CASME-II datasets is improved by 2.44%,2.64%,and 1.57%,respectively,compared to 3D-Dense Net.3.Use the trained micro-expression recognition model to construct a virtual environment interaction system based on micro-expression recognition for real-time face detection and micro-expression recognition.