Design And Implementation Of Micro-Expression Feature Extraction Algorithm Based On Deep Learning

Domestic and international extreme behavior,endangering public safety incidents are on the rise,such as network rumors,bus arson and driving collision sensitive areas,etc.On the early warning of dangerous behavior,relevant organizations and personnel began to study the early warning technology.Expression is an important nonverbal behavior of human emotion,which can be used as an important clue in the early warning of dangerous behavior.At present,some achievements have been made in the study of facial expression,but more attention is paid to the general expression.In addition to the ordinary expression,there are still difficult to detect the micro expression,its duration is very short,and potential intentions are closely related,this expression is a micro expression.For micro expression feature extraction is an interdisciplinary subject,involving multiple disciplines of computer,signal and information processing and clinical psychology,has important theoretical significance and practical application.It can help to promote mutual exchanges in various fields of study and promotion of related technology.This thesis focuses on the research of deep learning based micro expression feature extraction algorithm with Arousal(the degree of emotional awakening is sleeping),Valence(positive or negative emotions),Expectation(emotion is amazing degree)and Power(the degree of control of one's emotions by external influences),respectively.Finally,the predicted values are processed by a one-dimensional median filter.The main work of this thesis includes:(1)This thesis presents a new algorithm for convolutional neural network based micro expression features extraction algorithm.By comparing with the traditional algorithm of feature extraction such as Histogram of Oriented Gradient(HOG)and local binary patterns(LBP),the convolutional neural network based method can concentrate on the receptive field of micro facial expression,such as eyes and mouth.The results demonstrate that Convolutional Neural Network(CNN)can learn the micro expression representation with high level features from the original data.In addition,the performance of the algorithm does not depend on the accurate face detection and localization process.(2)This thesis presents a method of deep learning based predicting and classifying the micro affective factors.The proposed algorithm uses Multilayer Perceptron(MLP)instead of CNN with the fully connected layers.The experimental results on the AVEC2012 micro expression database show that the average recognition rate in four attributes(Arousal,Valence,Expectation and Power)are 71.51%,73.14%,66.43%and 69.05%,respectively.