Research On Facial Expression Recognition Technology Based On Convolutional Neural Network

Expression recognition as an important research direction of human-computer interaction applications,has been widely used in security,education,medical assistance,entertainment and other fields.In real life environment,facial expression recognition will be affected by factors such as illumination,facial occlusion,facial posture,etc.These complex noise variables make facial expression recognition technology still a challenging research topic.The existing facial expression image recognition algorithm based on shallow convolutional neural network has the defects of simple network structure,incomplete feature extraction,inadequate feature expression ability,and unsatisfactory recognition accuracy.Based on the research of convolutional neural network algorithm based on deep learning,this paper proposes an improved facial expression recognition algorithm based on convolutional neural network.This paper mainly conducts in-depth research on the relevant parts of feature extraction in convolutional neural networks.The main research contents of the paper are as follows:(1)Analyzed the problems of high data requirements,low accuracy and long training time when AlexNet is applied to facial expression recognition.This paper makes corresponding improvements to the AlexNet network structure,including five parts: continuous convolution,multi-scale convolution,Cross-connection and global average pooling and batch normalization make the network more suitable for expression recognition.And expand the experimental data set with a small number of pictures.The experimental results show that the improved AlexNet has achieved excellent results in facial expression recognition.Compared with AlexNet,the accuracy rates on the CK+,JAFFE and FER-2013 data sets have been improved by18.1%?6.08% and 29.89%,and There is a certain increase in training speed.The same improvement was carried out on the VGG network.The improved VGG network improved 9.96%?7.81% and 12.15% on the three data sets,respectively,which proved the practicability of this improved method.(2)The features extracted by the shallow network are not very expressive,and the amplified experimental data has a certain rotation.To solve this problem,a convolutional neural network combined with local binary pattern is designed for facial expression recognition,and the LBP feature map of the expression image is used as the input data of the convolutional neural network.The LBP feature map not only has rotation invariance but also the expression features of the image are more obvious,and the expression features of the facial expression extracted by the shallow network structure will be stronger.The experimental results show that the accuracy of the expression recognition algorithm combined with the LBP feature map on the three data sets is 92.34%,95.13% and 63.0%,respectively.The improved VGG network has accuracy rates of 90.55%,96.11 and 62.61% on the three data sets.(3)The single feature has the problem that the expression information is not comprehensive enough.Based on the idea of feature maps and convolutional neural networks,this paper designs LBP(local binary mode)feature map,RPCA(robust principal component analysis)Low-rank image combined with a dual-channel convolutional neural network facial expression recognition method.Two channels extract two facial expression features to complement each other,making the network more comprehensive and accurate for facial expression image learning and classification.The algorithm can achieve 97.89%,98.32% and 75.83% accuracy on three data sets.