Facial Expression Recognition Based On Deep Learning

Facial expression recognition (FER) is a currently active research topic in the fields ofcomputer vision, pattern recognition, artificial intelligence, etc. FER is an important part of theintelligent technology for human machine interaction，and recently it has drawn much attentionand numerous new methods have been proposed by the researchers in different fields. In thisthesis, we present the latest development of this area, and give a detailed analysis and summaryfor facial expression feature extraction and facial expression classification, which are the keytechnology in a FER system. At last, the current situation of FER are presented the futuredevelopment direction of FER.This thesis mainly studies some key issues in facial expression recognition featureextraction and classification, and puts forward some improvement methods by combining withthe method of deep learning. The main works of this thesis are as follows:1. Deep belief networks (DBNs) are a representative method of currently newly-emergeddeep learning theory. DBNs is capable of performing unsupervised feature learning, but couldnot perform classification. In order to effectively promote the performance of facial expressionrecognition, we propose a new method of facial expression recognition based on deep beliefnetworks. First, deep belief networks are used to learn the extracted primitive facial expressionfeatures, and get a higher level of abstract features used to initialize the hidden layer weights ofthe traditional model of multi-layer perceptron (MLP). Then we use the initialized MLP toperform the classification of facial expression. Experimental results on the JAFFE database andCohn-Kanade database, show that the proposed method can obtain the best accuracy of90.95%and98.57%for facial expression recognition, significantly outperforming the other usedclassification algorithms. It can be seen that the proposed method can be used to clearly improverecognition performance of facial expression recognition.2. A study on robust facial expression recognition based on deep belief network is given.This thesis focuses on investigating the performance of robust facial expression recognitionbased on deep belief based in the presence of corrupted expression images. Deep belief networkhas very strong ability of unsupervised learning, and can still achieve good recognition resultsunder different corrosion. Experimental results on the Cohn-Kanade database show that DBNshas excellent classification performance and robustness, and it is very suitable for facialexpression recognition. 3. The design of GUI interface of facial expression recognition is presented. Aftercompleting the process design of facial expression recognition, according to the basic principlesof the GUI system design—simplicity, consistency, constant learning, we have designed GUIinterface of facial expression recognition.