Research On Facial Expression Recognition Method Using The Deep Learning

With the development of intelligent machines and artificial intelligence,emotion analysis by computer will become more and more important.The investigation of facial expression recognition with emotional information application,has broad application potential in the field,such as teaching cognitive status analysis,psychological status of patients in public areas,warning of the danger,and visual perception of blind.In recent years,facial expression recognition,as the key technology of intelligent interaction and affective computing,has become a hot research topic in artificial intelligence in recent years.Considering that facial expressions have the characteristics of high-dimensional,non-rigid,multi-scale and vulnerable to illumination and angle,the analysis of accurateemotion change of the shape from facial expression or video by computer becomes difficult.At present,many researchers at home and abroad have researched on feature extraction and classifier design,and proposed methods based on global features,local features,gradient features,template features and so on.So far,these methods and techniques applied to all kinds of complex natural scenes still face many problems:1.The existing shallow feature extraction model effectively solves the problem of image dimension disaster,but leads to the effective loss of feature identification information,which restricts the accuracy of recognition.Furthermore,the algorithm promotion effect is limited due to multi-feature fusion,multi-scale extraction,multi-classifier combination.2.The fluctuations of algorithm recognition is obvious and robustness is low because of the complexity of the application scene,the face image vulnerable to angle,attitude,light,occlusion,multi-scale and other factors.Scene diversity and large sample data characteristics are required to establish the recognition model of complex space and to update the knowledge of posterior data.3.The existing static image algorithm applied to the natural scene make little use of the dynamic sequence information,which result in poor robustness of the algorithm.To date,the application effect to be improved because of less dynamic expression recognition model,the complex algorithm is,and the more assumption.In this paper,we research facial expression recognition methods based on the deep learningand explore autonomous learning model of visual features,provide theoretical analysis and technical support to build a more effective,end-to-end visual characteristics application to improve the accuracy and robustness of facial expression recognition.The main research contents of this paper are as follows:1.A facial expression recognition method based on the variation inference network is proposed.Considering that the shallow model is difficult to improve the accuracy of identification information,the variation inference network applied to the face recognition was designed.This method combining the advantages of self-coding network hierarchical feature learning and large data sample nonlinear fitting,variation reasoning algorithm generates the category distribution model of expression images.The mean value and the variance of the latent variables of the network can be obtained by inputting the measured expression data.Then,the network is fine-tuned according to the hidden variables,forming a stable probability generation identification network.The experimental results show that the proposed model can effectively perform hierarchical feature learning and form a complex nonlinear classification network.Compared with the self-coding network,the denoised self-coding network is faster and more accurate.2.A facial expression recognition method based on deep confidence network is proposedIn view of that the shallow model is difficult to improve the accuracy of identification information.This method combining the features of face region distribution can be identify facial expressionthrough the deep restricted Boltzmann machine(RBM)for data modeling and reasoning and integration of local and global features.Firstly,the hierarchical feature automatic learning structure is formed by the deep RBM,Then,according to the posterior sample combing the regional and facial expressions as well as the local and global features,a hybrid generating model has been formed.The experimental results show that the proposed method fits the facial expression distribution feature in view of the local feature and the global feature by the depth model to improve the accuracy of the recognition algorithm.3.A facial expression recognition method based on depth residual network is proposed.The method is aiming at the stability and convergence problem of the large-scale convolution neural network with the increasing depth of the network.As this method combines the network construction ability of the residual unit and the image feature learning ability of the convolution unit,adding the data sample to train the deep network,the complex discriminant model has been formed.Specifically,the input image is segmented into the parallel convolution network,the primary feature is extracted,then,the residual unit is used to superimpose the deep network for the deep learning of the aggregated feature.Then,to facilitate the integration of multi-scale features,each residual unit uses a multi-channel method for residual learning.Finally,the depth residual network output characteristics is coped with classification learning.The experimental results show that using residual learning to construct deep recognition network has high recognition accuracy and robustness.4.A dynamic facial expression recognition method based on LSTM+RNN is proposed.The method is aiming at the problem that the existing algorithms lack efficient use of the expression image sequence information and the robustness of the algorithm.We use recurrent neural network(RNN)to collect the image sequence,make the LSTM learning and the memory sequence correlate the information,and combine the single image information and the sequence related information to judge the expression.Thus,we can make the small-scale image data effectively locate the position of the face in the image,and then take advantage of the convolution of neural network to achieve visual feature extraction.Then,the overall structure of the cycle network can be built using RNN,which indicates the image sequence data association characteristics of a single LSTM unitlearning.Experimental results show that the method can obtain the relevance of the image sequence.Furthermore,accuracy and robustness have increased in view of the current image information for category identification.In conclusion,we explored the end-to-end expression recognition based on deep learning model.Better differential visual characteristics and nonlinear features of images have been obtained by learning the expression of image data automatically,hierarchically and effectively.These methods avoid experience needs and design flaws of artificial features extraction,which is an important research direction of the application of visual features.