| When people cover up their inner thoughts or feelings,their faces will involuntarily display subtle expressions that are hard to see with our naked eyes.Micro-expression is a type of expression with a small action range,weak performance intensity,and a very short duration(1/25s-1/2s).And it is extremely critical in the field of emotion monitoring,criminal detection,and national security.At present,most of the traditional micro-expression recognition models have low accuracies,the number of micro-expression samples in micro-expression datasets is relatively few,and the distribution of micro-expression categories is uneven.Therefore,how to accurately recognize and distinguish the categories of micro-expressions in a fleeting time has become a hot issue in current research.The work done in this paper mainly includes the following three aspects:(1)Data augmented for micro-expression dataset CASME II was developed.Based on the micro-expression dataset CASME II,the face position was accurately located by using the face key point detection technology in the Dlib library.After extracting the face image sequence,the corresponding keyframe image sequence was obtained by using the Temporal Difference method.Then,the keyframe images were augmented and balanced by using rotation,translation,and other technologies,so as to obtain the dataset CASME II+ that meets the deep micro-expression recognition model.This method not only reduces the interference of regional information outside the face,but also preserves the gradual process of facial micro-expression,and meanwhile expands the sample data size and reduces the imbalance of micro-expression sample distribution.(2)Micro-expression recognition based on the deep forest—DFMRM model was proposed.First,a multi-grained scanning module was utilized to extract micro-expression image feature vector,and then combined class vectors generated by random forests in order to get the transformed feature vectors,thus increasing the number of the feature vector and making the DFMRM model can be applied to the small training dataset.Then the cascade forest modules were employed to classify micro-expressions,the adaptive advantage of this module make training parameters of DFMRM are greatly reduced.Results show that DFMRM improves the accuracy of micro-expression recognition.DFMRM extends the application scope of deep forest and provides a new perspective different from the deep neural network.(3)Micro-expression recognition based on convolutional neural network—CNNMRM model was proposed.Firstly,the augmented micro-expression dataset CASME II+ was normalized to300×300 pixels;then,it is input into the hidden layer where the convolutional layer and the maximum pooling layer intersect each other for feature extraction,which effectively reduces the overfitting and effectively reduces the dimension of the feature.Then,the flatten layer is used to reduce the dimension of features,the dropout layer is used to enhance the orthogonality between each layer,and the softmax layer is used to complete the task of micro expression recognition.During the training,L1＿L2 regularization is added to the loss function to constrain the weight,reduce the overfitting,and improve the recognition accuracy,which can provide theoretical support for the application of micro-expression recognition.In summary,in this paper,the micro-expression recognition problems are studied,aiming at the problem of small and uneven distribution of datasets,and the overall accuracy is generally not high,thus this article enhances the dataset CASME II and gets CASME II + dataset,micro expression recognition models DFMRM and CNNMRM were proposed based on the deep forest and convolution neural network,respectively.It has been verified that the recognition accuracy of DFRMM model and CNNMRM model have achieved good results,reaching 98.03% and 98.44%accuracy,respectively.The method provided in this paper solves the problem of data imbalance between expressions,improves the accuracy of micro-expression recognition,and has certain reference significance in the field of micro-expression recognition. |
