Micro Expression Recognition Based On Convolutional Neural Network

Micro expression is a kind of spontaneous expression,which is produced when people try to hide their true feelings.It can neither be forged nor suppressed,so it can be used to estimate a person's real psychological state.Micro expression recognition has important application value in clinical medicine,criminal investigation,public security and other fields.In this paper,the micro expression recognition based on convolutional neural network is studied and analyzed and three improved micro expression recognition methods based on convolutional neural network are proposed.The main work of this paper is as follows:(1)A micro expression recognition method based on 2D-3D two stream CNN is proposed.In order to solve the problem that 3D convolution can't optimize time information and spatial information separately and simplify the scale of extracting time information,a two time scale three-dimensional convolution neural network is proposed.The spatial pooling layer is removed to improve the extraction ability of static expression.The parallel connection of temporal convolution and spatial convolution makes the parameter optimization process separate.Time convolution of different dimensions is added to increase the diversity of temporal information,so as to extract short-term features of micro expression.In order to obtain the motion information in the micro expression video,the motion information of the video is concentrated in one image by calculating the differential information image.The two-dimensional residual network is used to analyze the differential information image,and the long-term features of micro expression are extracted.In order to utilize both long-term and short-term features,a 2D-3D two stream convolutional neural network is obtained by fusion of the two time scale three-dimensional convolutional neural network and the residual network based on differential information image and to recognition micro expression.(2)A micro expression recognition method based on apex frame and motion region is proposed.In order to solve the problem of excessive redundant information in micro expression video,the cosine similarity between the current frame and the starting frame is used to locate the apex frame and remove the redundant video frame and the motion region in the video frame is captured to remove the redundant static features.In the apex frame,there is no temporal information between adjacent video frames,while the motion region in the video lacks complete spatial information.In order to extract the complete spatiotemporal information,a convolutional neural network based on apex frame and motion region is proposed and the features of apex frame and motion region are added to fuse and recognitions micro expression.(3)A micro expression recognition method based on attention mechanism is proposed.Aiming at the problem that there is less useful information in micro expression images and different regions have different effects on classification results,a convolution neural network based on spatial attention and channel attention mechanism is proposed.Different weight coefficients are distributed for different space or channel features to enhance the high correlation area and suppress the low correlation area.