Research On Micro-expression Recognition Based On Deep Learning In Video

The human face shows thousands of expressions through a special combination of muscle groups.Micro-expressions can directly reflect people's true emotions and psychological activities.It has a wide range of applications in customer interest discovery,public safety,psychotherapy,and business negotiation.Because micro-expressions are transient and subtle,the muscle groups that produce changes occupy a relatively small proportion in the image.Conventional expression processing methods cannot solve problems such as insufficient features,insufficient samples,and insufficient real-time in micro-expression recognition.At present,deep neural network-based methods are easier to meet the real-time application than traditional feature extraction methods.The quality of the extracted features becomes the decisive factor for the accuracy of the model recognition.Based on the above problems,this research is based on deep neural networks,and conducts research from the perspective of efficient extraction of image features,construction of end-to-end networks,and optimization of model training processes.The micro-expression data is divided into 7 types,and the specific research content and contributions are summarized as follows:(1)For the problem that information loss and insufficient ability of feature vector expression caused by convolution operation,a lightweight CNN micro-expression recognition method based on optical flow features is proposed.This method completes image pre-processing through segmentation and cropping,so as to avoid extracting spatio-temporal features of unrelated changes in the background;The LiteFlowNet is used to calculate optical flow and construct an effective feature map that fuses optical flow,texture,and grayscale features to supplement the lost detailed features of the convolutional network.On this basis,the feature classification is completed,and the accuracy rate can reach 60%.(2)When multiple convolutional layers are superimposed on the micro-expression recognition network,the low-frequency information of the image will be lost.In order to solve this problem,this paper proposes a method based on dilated convolution microexpression recognition.In order to improve the details and fully extract the microexpression features,the number of dilated convolution layers and the corresponding dilated convolution rate are determined in the article by comparing the effects of different dilated convolution rates on model recognition accuracy,the CNN network feature extraction process is improved.The model introduces a face automatic correction module to improve the multi-angle image input problem in practical applications,meets the model's real-time recognition and classification needs,and completes model evaluation by comparing different optical flow effects.(3)Because the data samples of micro-expressions are difficult to expand.In order to speed up the model training speed,this paper proposes a micro-expression recognition model combining dilated convolution and residuals.The model solves the vanishing gradient problem through the residual block,and accelerates the model training convergence rate.Aiming at the imbalance problem of network training samples,different loss function schemes are used to control the balance of positive and negative samples.The average accuracy is 72.26%,and the real-time performance of video analysis is maintained at 60 fps.