Micro-expression Recognition Based On Deep Convolutional Neural Network

In recent years,under the influence of the continuous upgrading of computer hardware technology and the arrival of big data and cloud computing technology,researchers in image processing and recognition have also targeted more subtle features in images.In addition,face recognition and micro-expression recognition have been widely used in many fields such as anti-terrorism,criminal investigation,and mental illness medical treatment.For this reason,micro-expression recognition is gradually becoming a research hotspot of researchers.Micro-expressions are generated by real emotions triggering the movement of human facial muscles,and are the embodiment of human real emotions.Due to the short duration of micro-expression,extreme features and high difficulty in data collection,and the public data set supporting its research is very rare at this stage,training a stable micro-expression recognition model is still a problem to be solved at this stage.problem.To this end,this article is based on deep convolutional neural network for model training,and the dimension is increased to three dimensions,which further improves the recognition performance,stability,and generalization ability of the model.The main work done in this paper is as follows:1.Aiming at the problem of difficulty in extracting micro-expression features between images with and without micro-expression in static images.This paper extracts the features of micro-expression from a three-dimensional image sequence,and uses a three-dimensional convolutional neural network to extract information in the time domain from a sequence of video frames,thereby facilitating the extraction of the action features of the micro-expression.The experimental results show that the average recognition rate of the model reached 76.0% under the casme? data set,and the recognition rate of the model reached 78.2% under the SAMM data.Compared with the existing methods,the recognition rate has been improved,which further proves The advantages of this model for micro-expression recognition,and when a different test set is changed,the recognition rate of this model is basically the same as the original test set,which further proves that the model has good stability and generalization ability.2.In response to the problem of the limited set of micro-expression data available in the world.This article uses the casme2 data set released by the team of Xiaolan Fu of the Chinese Academy of Sciences and the SAMM data set of 32 participants from13 different races from abroad as training samples.In the preprocessing stage,this article stores the pictures in CSV format,and then formats them as.h5 standard database files,which further improves the efficiency of data processing and uses the Dropout method to prevent overfitting.Finally,the softmax method is applied to classify the final output vector.The simulation shows that this method reduces the training time,reduces the hardware pressure,reduces the negative impact caused by the small data set,and improves the generalization ability and robustness of the network.3.Aiming at the problem of low recognition rate and low learning ability of the micro-expression recognition model,this paper proposes a micro-expression recognition method based on transfer learning.Feature extraction is performed by extracting five features of three-dimensional image sequences,and then the two convolutional neural network structures of Res Net34 and VGG19 are transferred for network training respectively,and finally the network is tested on the casme? data set.Simulation experiments show that the recognition rate of the two networks is 78.9%and 77.9%,and the recognition rate is better than similar machine learning algorithms.The trained model has stable learning ability,which improves the model recognition rate,which shows the feasibility and efficiency of this method.