Research On Facial Expression Recognition Method Based On Improved Convolutional Neural Network

Facial expressions contain rich emotional information,which is an important way for humans to transmit information,it can intuitively reflect human inner thoughts and plays an indispensable role in the process of interpersonal communication.With the vigorous development of computer vision and artificial intelligence,facial expression recognition technology has attracted widespread attention of scientists and has gradually become a research hotspot.At present,facial expression recognition technology has been widely used in intelligent transportation,medical care,humancomputer interaction,criminal investigation,online education and other fields,which has very important research value.Facial expression recognition technology classifies the facial expression features by extracting the facial expression features,so as to determine the category of the facial expression,and the key to the facial expression recognition technology is how to effectively extract the facial expression features.Traditional facial expression recognition methods mainly rely on manual methods to extract facial expression features,which have shortcomings such as low feature extraction efficiency,complex feature extraction process,and low recognition accuracy.With the rapid development of deep learning and convolutional neural networks,facial expression recognition technology has also been rapidly improved.Convolutional neural networks can automatically extract facial expression features,but how to design a convolutional neural network so that it can extract more comprehensive and differentiated features,thereby improving the accuracy of facial expression recognition of convolutional neural networks,is still It is the focus of facial expression recognition research.In response to these problems,this paper selects the VGG16 network and Res Net18 network as the basic network for facial expression feature extraction,and improves on this basis,thereby improving the recognition accuracy of facial expression.The main research work completed in this paper is as follows:1.According to the development trend of facial expression recognition and the subject requirements,on the basis of mastering the relevant theories of deep learning,convolutional neural network and facial expression recognition,the overall design scheme of facial expression recognition in this paper is proposed.2.To solve the problems of insufficient facial expression feature extraction and low recognition accuracy in the VGG16 network,this paper proposes an improved facial expression recognition network model based on the VGG16 network.Firstly,the VGG16 network is divided into five blocks,and then the features extracted from the last three blocks of the VGG16 network are fused,and the shallow features and deep features are fused to make the feature information extracted by the network more sufficient.And add the SGE(Spatial Group Enhance,SGE)attention module to help the network extract facial expression features more accurately.At the same time,the fully connected layer is deleted,and a fully connected layer is used to directly output the classification result,reducing the amount of parameters of the network.The experimental results show that the recognition accuracy rates of the original VGG16 network on the FERPlus dataset、RAF-DB dataset and SFEW dataset are 87.41%、81.35% and 50.00%respectively.The recognition accuracy rates of the improved VGG16 network in this paper on the FERPlus dataset、RAF-DB dataset and SFEW dataset reached 89.50%、86.70% and 56.88%,respectively.Compared with the original VGG16 network,the recognition accuracy rates were increased by 2.09%、5.35% and 6.88%,respectively,which shows that the improved VGG16 network in this paper has higher facial expression recognition accuracy.3.To solve the problem that the Res Net18 network has feature information loss when downsampling and inadequate feature extraction,which leads to low recognition accuracy,this paper proposes an improved facial expression recognition network model based on the Res Net18 network.Firstly,the residual block and the downsampling layer of the Res Net18 network are improved to mitigate the problem of feature information loss during downsampling by using a Local Importance-based Pooling.Then add a Style-based Recalibration Module to recalibrate the channel weights of the output feature map to improve the network’s ability to represent important feature information.Finally,the extracted features are sent to the fully connected layer for facial expression classification.The experimental results show that the recognition accuracy of the original Res Net18 network on the FERPlus dataset and the FER2013 dataset are 87.83% and69.99%,respectively.The recognition accuracy of the improved Res Net18 network in this paper on the FERPlus dataset and the FER2013 dataset has reached 89.75% and 73.95%,respectively.Compared with the original Res Net18 network,the recognition accuracy has increased by 1.92%and 3.96%,respectively,which shows that the improved Res Net18 network in this paper has higher facial expression recognition accuracy.At the same time,based on the improved Res Net18 network in this paper,a facial expression recognition system is designed,which can realize static images,dynamic video and real-time facial expression recognition.