Research On Micro-expression Recognition Based On 3D-CNN

Facial micro-expressions refer to transient muscle changes in the face,indicating that a person is consciously or unconsciously suppressing his true emotions and even mental health.Therefore,micro-expression recognition has attracted more and more research work in the field of psychology and computer vision.It can be used in psychological research as the theoretical basis of polygraph detection,and also in clinical diagnosis,criminal interrogation,transaction negotiation Play a role.The research on micro-expression recognition has been in its infancy.The main problem is that due to the small pressure and short duration of micro-expression facial expressions,micro-expression recognition has become a challenging task.Identification research brings some difficulties.With the development of deep learning,computer vision has become a hot research field in recent years,and various theories and technologies have emerged one after another,which has played a good role in promoting micro-expression recognition.The deep convolutional neural system has proved to be highly effective for difficult facial recognition tasks.On this basis,this article explores the extension of the convolutional neural network model to a 3D convolutional neural network.The 3D model extracts the dynamic expression information in the video,trains the proposed model on the micro-expression dataset,obtains the accuracy of the model,and compares it with the traditional method to show the experimental results.The research work includes the following three aspects:(1)For the dynamic micro-expression sequences in the video,it is proposed to use the 3D convolutional neural network model to extract features.One of the structures is a 3D convolutional neural network model that can process RGB streams in the video.The facial contours detected in the video are treated as For recognition object processing,the network structure is simple;the other is to use the GoogLeNet network architecture as the model basis,extending the Inception substructure from the spatial domain to the space-time threedimensional domain to process dynamic video,and as the simultaneous processing of RGB streams and light in the video Streaming dual stream 3D convolutional neural network.(2)Train the proposed 3D convolutional network model on the Tensor Flow platform,use GPU parallel computing to improve the efficiency of the experiment,the data set used for the experiment is and two micro-expression data sets,and the test results and the methods over the years Through comparative analysis,relatively good test accuracy is achieved.(3)Finally,build a Web-based micro-expression recognition service application to give classification accuracy measurement for the micro-expressions in the input video.The video stream information is obtained through the audio and video engine in WebRTC technology,and the effective information in the video stream is obtained through the Kurento face detection module and filter module.Finally,the overall architecture and process of the system are introduced.