Research On Face Expression Recognition Based On Convolutional Neural Network

Facial expressions play an important and fundamental role in interpersonal and social interactions,revealing all kinds of information,such as emotion,identity,gender and age.In the past few decades,automatic recognition of facial expressions has become an active research.The key to facial expression recognition is the accuracy of feature point detection,the quality of feature point detection will have an important impact on the classification of the system.Convolutional neural network is one of the most advantageous methods in feature point extraction in recent years.Research on facial expression recognition based on convolutional neural network has important theoretical significance and application value.The overall method of automatic facial expression recognition is usually to locate a face for a given input image,detect a set of face feature points,and then use these points to represent the face.The final step is to use the extracted feature vector and use a classifier perform expression recognition tasks.This thesis starts from the face feature point detection based on the traditional method,and further studies the key point detection algorithm based on the convolutional neural network.On this basis,two facial expression recognition algorithms are proposed.The main work is as follows:(1)Researched the traditional points face detection and expression recognition method and the method based on convolutional neural network,detailed analysis of the research route of different methods and the technical difficulties faced by them,and a detailed comparison of the advantages and disadvantages between the two methods.(2)A face feature detection method based on deep learning is studied,in order to reduce the information redundancy between channels and pay attention to the part with the representative information in the spatial feature map,a convolutional neural network based on an improved ResNet residual module is proposed to extract facial features.The residual network introduces a channel attention module and a spatial attention module,which adaptively aggregate feature maps in the channel and space domains to learn the relationship matrix between channels and the relationship matrix between spaces.The proposed method is verified on the 300 W face public dataset,and the best results are obtained,which provides better feature descriptors for the facial expression recognition algorithm.(3)On the basis of obtaining face feature points,this thesis proposes two different recognition schemes for the task of face expression analysis.By analyzing the different feature points extracted by SDM,ResNet and improved ResNet,this thesis proposes a face expression recognition algorithm based on multi-feature fusion,and conducts experimental verification on the public data set.However,the experimental results still do not reach the optimal performance,there is still room for improvement.So more in-depth study and put forward a kind of face facial expression recognition based on ensemble learning algorithm,this method not only avoids the feature fusion method of feature vectors between homogeneous enhancement,heterogeneity weakens,but also fully take advantage of all the feature descriptor(including low-level features and high-level),in the experimental section,Compared with the recognition results of several advanced neural networks on the public data set,the algorithm in the thesis has achieved the best results.