CNN-RNN Based Micro-expression Recognition

Human facial expression is an important information of human expression,and the recognition of facial expressions can promote the understanding of human psychological state.Compared to the average expression,it lasts very short,usually only 1/25 to a third of a second,and the range of motion is very small and extremely difficult to find.Micro expression is a kind of can not inhibit the spontaneous expression,also cannot hide so judge people can be used as an important basis of true feelings in the psychological consultation and the national security and other fields have a wide range of applications.Micro-expressions artificial recognition need related trained professionals and identification accuracy is not satisfactory,therefore micro-expressions automatic identification has become one of hotspots in the field of computer vision in recent years.But because currently use more traditional micro-expressions automatic identification method requires manual in advance to extraction of shallow characteristics,the consumption of resources,and points out some problems such as lack of feature extraction,lead to the recognition effect is not ideal.Deep learning in recent years because of its prominent feature extraction and classification ability has made breakthrough progress in the field of image recognition,so scholars have already begun to deep learning method is applied to micro expression recognition problem and achieved better effect than the traditional method,but there are still room to improve recognition rate.In order to further enhance the performance of the recognition algorithm,this paper proposes an improved deep learning model of micro expression of spatial and temporal characteristics of feature extraction,micro expression recognition task better.The main contents of this subject are as follows:(1)For sample data set between frames are not uniform,to improve the traditional frames normalization method,put forward a number of frames based on convolutional neural network normalization method,and complete the amplification and handling of data sets.Previous micro-expression recognition method using only a single time interpolation model(TIM)on the micro expression sequence frames and unified samples,when will this lead to large difference between two frames with the target sequence actual frames when the frame is obtained by frames unified image characteristics losses significantly,is not conducive to deep learning network to extract the features,and then affect the identification accuracy.This paper proposes a frame based on convolution neural network into the frame number unified method,with the method of the data set all micro-expressions sequence frame processing,reoccupy TIM model frames normalized in order to reduce the normalized micro-expressions characteristics in the process of the loss.In this paper,the data set is amplified to prevent the over-fitting phenomenon during the process of the data set.The experimental results show that the improved frame number normalization method can achieve better results,and the amplification of the data set can significantly reduce the over-fitting phenomenon.(2)On the depth of the current application more widely,comparing the learning network needs to choose to facial expression recognition the task of network model,and on the basis of improvement,to characteristics of micro expression sequence space sequence characteristics were extracted in time.In this paper with a single convolution neural network for feature extraction of micro-expressions series can extract to micro-expressions space characteristics and ignored its temporal characteristics of the problem,the convolutional neural network with the strong ability to deal with timing characteristics of circular neural network combination,make it the spatial and temporal characteristics of micro expression is extracted,better classification and recognition.In view of the image sequence processing deep learning approach training for a long time,large computational complexity,algorithm,from the network layer,the number of parameters,accuracy comparison of the current commonly used several kinds of CNN model and RNN model,finally set up according to the requirements of micro expression recognition task CNN micro-expressions-RNN model identification.Micro-expressions recognition method proposed in this paper the CASME Ⅱpublic micro-expressions datasets achieves a recognition rate of 69.82%,achieved better effect than ever before.