Research On Micro-expression Detection And Active Learning

Micro-expressions are brief facial movements characterized by short duration,small change amplitude,small motion range and involuntariness.Due to the special psychological mechanism of micro-expressions,their appearance often means that people are trying to conceal their true emotions.Therefore,studying micro-expressions has great practical significance in the advancement of psychology.For example,micro-expressions are generally known for their potential use in national security,clinical diagnosis and case detection.However,it is difficult for naked human eye to detect micro-expressions due to the short duration and low intensity.With the related research in the field of computer vision and pattern recognition,many achievements have been made in micro-expression detection and recognition.Existing detection algorithms where micro-expression clips detected are discontinuous have some shortcomings and their detection effect is not ideal in micro-expression detection.In addition,the collection and calibration of micro-expressions are difficult,time-consuming and labor-intensive with high cost,which leads to insufficient sample size of the micro-expression database and limits the research of micro-expression recognition algorithm.Therefore,it seems especially important to improve the performance of micro-expression recognition with a relatively small sample size.Aiming at these problems above,this paper mainly studies the micro-expression detection and active learning.In micro-expression detection,a micro-expression detection algorithm based on the combination of modulus and angles of optical flow feature vector in perceptual region is proposed.At the same time,in micro-expression recognition,two micro-expression recognition methods based on active leaning are proposed.The main work and innovations of this paper are as follows:First,micro-expression detection detects micro-expressions in the video and locates the start frame,the climax frame,and the end frame.Therefore,this paper proposes a micro-expression detection algorithm based on the combination of modulus and angles of optical flow feature vector in perceptual region.According to the distinction of motion characteristics of different expressions' FACS motion units,we present a dividing method based on perceptual region in human faces.A method of combining the optical flow modulus and the angle is proposed,according to these perceptual regions,the whole face and all the local AU motion units are synthesized,and a new decision criterion is proposed to detect the micro-expression segment visually,which greatly improves the effect of micro-expression detection.Second,due to the collection and calibration of micro-expressions are difficult,time-consuming and labor-intensive with high cost,the sample size of the micro-expression database is insufficient,which leads to poor performance in micro-expression recognition,limits the automatic recognition research and its application.Hence,this paper proposed two micro-expression recognition algorithms based on active learning.A:A micro-expression recognition method based on information and representative active learning.By finding the sample that minimizes the objective function,we can get a sample that provides both information and representativeness.A systematic approach to measuring the accuracy and representativeness of information is presented based on the minimum and maximum view in active learning.This method facilitates the selection of valid samples and get a more efficient data set with the lowest cost.B:A micro-expression recognition method based on edge probability distribution matching for batch mode active learning.The purpose is to ensure that the classifier learns on the marked data with similar distribution which can represent unmarked data,so that the unmarked data and unlabeled data from the same distribution will have favorable generalization performance.Using this method,we can query a specified number of samples at a time.and the information provided between the samples is comprehensive.which can avoid the repeated iteration process,reduce time consumption and redundant information and make full use of resources,especially when there is a large volume of data.