Research On Facial Micro-expression Recognition Based On Machine Learning

Expression is an important nonverbal behavior of human beings to express emotions.Irrepressible and forged micro-expressions can provide more real information than macro-expressions.Micro-expressions are so short in duration that they are almost impossible to be captured with the naked eye.Using this feature,they could be used in lie detection,psychological counseling,national security,and even new types of classroom education.In daily life,it is very difficult to judge micro-expressions manually unless you are professionally trained,and a psychologist who specialized in recognizing mic'ro-expressions is also difficult to be trained.Therefore,how to use machines to automatically recognize micro-expressions has become a hot research direction.Most traditional automatic micro-expression recognition methods need to manually extract the superficial features of face in advance,which will increase a lot of work,and exist some image information loss.As a result,the recognition accuracy can not reach the expected target.In recent years,machine learning has been applied in the field of image classification and recognition by more and more scholars due to its excellent feature extraction and classification ability.As one of the research directions,micro-expression recognition has achieved good results in accuracy compared with the traditional classification and recognition methods.However,machine learning is also accompanied by parameter explosion,overfitting,high hardware requirements,slow model training,failure to guarantee real-time performance and so on.Aiming at these points,a two-dimensional stochastic configuration network has been proposed,which can directly process two-dimensional image data,to classify and recognize micro-expressions by improving the stochastic configuration network originally used for classification and processing one-dimensional data.It not only guarantees the accuracy,but also reduces the requirement of hardware and speeds up model training.At present,the open micro-expression database is not complete,the number of micro-expression samples is not large,and the problem of non-uniform face position in the samples still exists.So,the original data set is preprocessed to complete face cutting,alignment,image normalization,data set expansion and other operations.By comparing several typical neural networks which are widely used in image processing at present,it is found that there are many problems in traditional deep neural networks,such as too many parameters,huge computation,complex network structure,being easy to produce overfitting,slow model training speed and so on.An improved two-dimensional stochastic configuration network based on incremental single hidden layer random weight feedforward neural network is proposed to classify and recognize micro-expressions.A two-dimensional stochastic configuration network which can directly input images and is suitable for micro-expression recognition is constructed.By experimental comparison with the basic stochastic configuration network and several typical deep neural networks,it shows that the two-dimensional stochastic configuration network has good generalization ability,acceptable effectiveness and sufficient efficiency in micro-expression recognition.