Research On Video Face Posture And Expression Based On CNN Logic Architecture Analysis

Recently years,with the vigorous development of artificial intelligence,pattern recognition and computer vision technology,the research of face information has gradually become a hot direction.Face,as an important biological feature,is widely used in many fields such as human-computer interaction,medical treatment,personal safety,daily life and military.The research of face information mainly focuses on the information of face posture,skin color and expression in video and image.At present,most of the research on face information mainly base on pictures,and almost all of the face information of a certain information research,two or more kinds of face information research is still relatively few.The rapid development of deep learning significantly improves the recognition rate of face information,however,the analysis of the logical structure of the network is rare.In this paper,we use convolutional neural network to study face information,and then analyze each network model.The main works of this paper are as follows:In the stage of facial expression recognition,this paper chooses a suitable convolutional neural network structure by observing visual feature maps and accurate rate polyline maps.Firstly,three network models,Alex Net,VGNet and Res Net,are used to train CK + and JAFFE datasets,and their network models,accuracy and error loss values are well preserved.The output features of each layer of network are visualized,and the accuracy and error loss polygons of the three networks are drawn at the same time.The most suitable network model for facial expression recognition is selected through analysis.Then a better network model is selected and trained,and the accuracy of the final test set reaches 98.3%.When classifying face posture,this paper proposes a face posture classification method based on transfer learning.Firstly,the convolution neural network is used to train one-way samples and preserve the network model.Then,when training two-way network models,firstly,the trained network model is loaded,and then the two-way samples are trained.The final accuracy rate reaches 98.7%.The network model is used for multi-pose face recognition.Before multi-pose face recognition,the network model using face pose is roughly classified into three categories,and then face recognition is performed,which achieves high accuracy.After obtaining the network model of multi-pose face recognition,three network models are visualized and compared with the features extracted in face pose classification.The relationship between the learning task and the features extracted from the network model is obtained.The output of the network model of face posture is analyzed,and the more active neurons in the network are obtained by feature matching method,and the statistics are carried out.This paper trains network models related to face information by convolutional neural network,and achieves high accuracy.Through the analysis of the output characteristics of each network model,the corresponding rules are obtained.