Micro Expression Recognition And Research Based On Deep Learning

Facial expression plays a very important role in human non-verbal communication,which is a very common way of emotional expression.Facial expression recognition can be applied in different situations.Here,facial expression generally refers to macro expression,that is,the expression with large action range and obvious regional characteristics.With the support of the current research technology,the recognition of macro expression has achieved ideal results,but there are some common problems in the field of expression recognition,such as image resolution,illumination,facial image occlusion and so on.The category of expression is more extensive.Facial expressions such as "happiness,anger,sadness and happiness" that can be clearly recognized by the naked eye are called macro expressions,which are also commonly referred to as expressions.The characteristics of micro expression are reflected in different aspects,such as action time,action range and uncontrollable spontaneity.The duration of the action is only 1 / 25 ? 1 / 5S,and the change range of the action is very weak,so it has become a major challenge in the field of expression recognition.Micro expression is often used in the filed investigation,medical health,social security and other fields,which has great application research value.Traditional machine learning methods usually use manual model definition,feature extraction and classifier establishment,which is often cumbersome.With the development of deep learning technology,this research can also be carried out in the field of micro expression.The main research directions of this paper are as follows:1)This paper studies the micro expression data set and preprocessing,analyzes the difficulties of micro expression database establishment and compares with the existing data set.The preprocessing work of micro expression data is introduced,and the data expansion work is carried out in CASME2 data set and SMIC data set.2)A micro expression recognition method based on ResNet50-LSTM is proposed for experiment.The 50 layer residual network structure is used to extract the spatial features of the data,and the deeper network layers are used to extract the effective features of the data.Aiming at the phenomenon that the information between adjacent frames in micro expression is correlated,the temporal network part designs long-term and short-term memory units for feature learning,and solves the problem in time domain in the process of micro expression recognition.Through the combination of the two methods,the effectiveness of feature extraction is greatly guaranteed.Compared with the traditional methods,the recognition accuracy has been improved.