Research On Stack Hybrid Autoencoder And Transfer Learning In Facial Expression Recognition

Face expression recognition,as a key technology in the intelligent system of emotional recognition,can well express human psychology,emotion,intention and so on.It is one of the important foundations of human-computer interaction.However,the traditional facial expression feature extraction method not only increases the training time and space complexity of the model,but also loses the expression feature information of the original image to a certain extent,which affects the recognition effect.Therefore,this thesis introduces the method of deep learning into the task of facial expression recognition,The main works of this thesis are as follows:1.In the process of facial expression recognition,the problem of extracting better facial expression features and improving recognition accuracy is discussed.In this thesis,denoising autoencoder,sparse autoencoder and ordinary autoencoder are cascaded to form a multi-layer fully connected network.Then a new classification network,stack hybrid autoencoder,is formed by adding softMax classifier to the last layer of the network.According to the greedy layer-by-layer training method(that is removing the output layer of the former autoencoder and using the hidden layer of the autoencoder as the input layer of the next autoencoder),the local optimum problem is solved,Because the net,work combines multiple autoencoders,the network has the advantages of multiple autoencoders.It makes the extracted features more representative.The stack hybrid autoencoder network is used to simulate on JAFFE(The Japanese Female Facial Expression Database)and CK+(The Extended Cohn-Kanade Dataset)databases,and the recognition accuracy is 96.38%and 96.70%respectively.2.Aiming at improving the accuracy of facial expression recognition in facial expression simulation experiments,when the facial images used in testing and training do not satisfy the same distribution.this thesis adopts a method of migrating convolution network.Firstly,a new full connection layer is defined,and then the convolution layer of the three neural networks(AlexNet,VGG16,Inception V3)which have achieved good results in the classification competition of ImageNet database is migrated to the newly defined full connection layer to form a model migration network for expression recognition.The experiment was carried out using a hybrid database composed of CK+ database and JAFFE database.The experimental results show that the three migration networks adopted in this paper have achieved good recognition results,among which Inception V3 migration network has the best recognition effect.It converges basically when the number of iterations is 800,and the recognition accuracy reaches 93.86%.