Research On Facial Expression Recognition Methods Based On Images

Facial expression recognition is one of the hot-spots in the field of emotional computing and human-computer interaction,and has broad application prospects.The facial expression recognition algorithms have achieved rapid development in recent years.The recognition algorithms based on traditional machine learning is generally aimed at a spot of data samples.When the feature dimensions are too more,it will cause over-fitting.When there are many data samples,the deep learning algorithms can be used to distinguish them.But it costs much time to train the model of the Neural Network.How to efficiently extract facial expression features and reduce the training time,so that further improve expression recognition rate is very important for us.In view of the above problems,this thesis has mainly completed the following work:(1)For the traditional machine learning methods,where there are a spot of data samples,it will produce the high dimension of feature,so we use the algorithm of LBP+KDIsomap is proposed to reduce the dimension.The algorithm of LBP is used to extract the features and the algorithm of KDIsomap is used to reduce the dimension of feature;In order to reduce the Misjudgment of samples,we use the SVM-KNN classifier to instead a single classifier.(2)In order to make full use of the temporal flow feature information of facial expression images,a facial expression recognition algorithm based on Two-Stream Convolutional Neural Network combined with Support Vector Machine is proposed.Two-Stream Convolutional Neural Network can automatically extract the geometric features,texture information and time-flow information of the expression images.What's more,it also utilizes the powerful nonlinear classification ability of SVM to make it possible to achieve higher expression recognition rate.(3)For the deep learning methods,the face expression recognition method based on the preliminary training Multiple Depth Fusion Convolutional Neural Network models is proposed to solve the problem that the training time is too much.By combining multiple deep convolutional neural network models to extract more rich facial expression feature information,it is beneficial to improve the recognition rate.At the same time,the Preliminary training model parameters can be migrated,which can reduce the training time of the networkmodel.CK+ database is used to test the proposed methods of this thesis.The experimental results show that the three methods proposed in this thesis can improve the expression recognition rate and achieve the expected purpose,and verify their effectiveness and feasibility.