Research On Micro-expression Recognition Based On Convolutional Neural Network

Facial expressions have been studied extensively for decades,both in psychology and pattern recognition.Micro-expression recognition is also on the rise,it has more challenge.Micro-expression is the expression of subtle emotions on the face.It is a rapid and weak facial movement that is almost beyond human control.Although great progress has been made in the study of facial micro-expressions in recent years,there is still much room for improvement in the recognition of micro-expressions.Micro-expression analysis has been applied in many fields such as emotion monitoring,criminal investigation and psychological therapy.Compared with traditional expression recognition,the recognition of micro-expressions is more challenging,because the duration of micro-expressions is very short and the facial movements involved are tiny.So far,the work in this field is still in its infancy,with only a few micro-expression databases and methods.In the early stage of micro-expression recognition,manual features are mostly used.With the development and progress of deep neural network,the research on micro-expression recognition using deep neural network becomes more meaningful.This paper studies two network models that work well on other computer vision issues,such as Inception and Resnet.It mainly includes: 1)In order to improve the recognition effect of micro-expressions,construct the TIncep SE network with added attention mechanism;2)In order to utilize the micro-expression sequence information,relevant experiments were carried out on 3DResnet series network with 10 frames of micro-expression picture sequences as input;And using TIncep SE and 3DResnet two network models to form a dual-stream network model 3DRes TIncep SE.Experiments show that the dual-stream network proposed in this paper has a good recognition effect on CASMEII,SAMM and SMIC2/HS.Among them,the best deep learning method is exceeded in SAMM and SMIC2/HS databases.In particular,the Accuracy of SAMM is improved by 17%,and the comprehensive evaluation index F1 is improved by 29%.