Research On Micro-expression Recognition Technology Based On Video Amplification

Micro-expression is an expression expressed by a person's facial muscles with extremely small movements,which usually occurs when people deliberately conceal and suppress their true feelings.Compared with macro expressions,micro expressions have the characteristics of short duration,small expression amplitude,and difficult to observe.Usually a micro expression has a duration of 1/25 seconds to 1/5 seconds,so it is difficult to capture.Because of the above characteristics,it poses a huge challenge to microexpression recognition.Even trained professionals are in a low level in judging microexpressions.With the rise of computer vision and the release of multiple spontaneous micro-expression data sets,micro-expression recognition algorithms have developed rapidly.The micro-emotion data has a small action range,and the recognition algorithm has higher requirements for data quality.Data quality is greatly affected by the acquisition environment and equipment.This paper introduces a dark channel-based denoising algorithm to improve data quality.It is theoretically proved that the micro-expression data meets the dark channel prior theory,and the dark-channel denoising algorithm is compared with various denoising algorithms applied to the micro-expression data.The experimental results show that the denoising algorithm based on dark channels has the most significant improvement in data quality.Motion video magnification technology is a technology that helps people observe fine movements after magnifying small changes in the video.In this paper,micro-expression features are magnified with the help of motion amplification algorithm,which is convenient for the recognition and classification of micro-expressions.There are two main theories of sports video magnification technology,Lagrangian theory based on prediction of motion behavior,and Euler theory based on frequency and time domain.Early motion video amplification algorithms were based on manually designed features and filters,while motion learning networks based on deep learning learned features and filters from convolutional neural networks autonomously.In this paper,theoretical analysis proves the feasibility of applying motion video amplification to micro-expression recognition.Then experimentally compare the effect of phase-based motion video magnification algorithm and deep learning-based motion video magnification algorithm on micro-expression motion magnification,and introduce structural similarity parameters to evaluate the image quality of the magnified data.In this paper,the CASME ? data set is used as the evaluation data set,and the leaveone-subject-out cross-validation method is used to calculate the model classification accuracy,recall rate,Macro-F1-score and other evaluation parameters.In order to analyze the impact of the motion video amplification network on the classification results,this paper calculates the classification evaluation parameters and confusion matrix of multiple motion amplification coefficients.The short-term memory network based on motion amplification proposed in this paper can effectively improve the misclassification problem caused by the one-to-one correspondence between micro-expressions and facial coding units.The experiment proved that Macro-F1-score achieved 57.42% in the crossvalidation of the MMSTM network proposed in this paper.