Multi-view Facial Expression Recognition Based On Feature Mapping

Facial expressions contain a wealth of emotional information.It is an important way of communication except language,words and voice.Now most research is aimed at frontal facial expression recognition.However,the frontal facial expression is idealized,because in real life,the acquisition device may capture the expression data from all angles.Therefore,for the multi-angle facial expression recognition,a more robust facial expression recognition algorithm is proposed which has been of great significance in the field of human-computer interaction.Due to the lack of multi-view facial expression database and the limitation of algorithm,the research on multi-view expression recognition is limited.The existing multi-view facial expression recognition method involves training view-specific classifiers or learning all angles with a single classifier.However,these methods ignore the fact that different angles of facial expression are only the different show of one same expression.Therefore,based on the study of frontal facial expression recognition,this paper proposes a method that using neural network to map multi-angle expression features into frontal facial expression features and then classify them.This algorithm solves the problem that time-consuming of training view-specific classifier,and improves the robustness of multi-angle expression recognition.The main work of this paper is as follows;Firstly,we introduce the research background and significance of facial expression recognition.The current status of multi-angle expression recognition is summarized from three aspects which is face detection,feature extraction and classification,and the existing human face expression recognition algorithm is analyzed,including frontal expression recognition and multi-view expression recognition.After comparing the advantages and disadvantages of every algorithm,feature mapping method is adopted to solve the multi-view expression recognition problem.This paper also carries out the experiment of face detection and image preprocessing.The feature of Pyramid histogram of oriented gradients is choosed as the feature of this paper,which is not sensitive to the direction and scaleSecondly,we achieve the feature mapping method for multi-view facial expression recognition.RBF neural network not only has a simple structure,but also has strong non-linear mapping ability and self-learning ability,especially local approximation ability and learning speed performance.Therefore,we choose the radial basis function neural network to carry out the mapping work from multi-angle expression feature to frontal expression featureThirdly,we design a multi-view facial expression recognition algorithm framework.Based on the extracted PHOG features,the nearest neighbor classifier and the sparse representation classifier are studied,and then we analysis the nearest neighbor sparse representation classifier.The experiment result shows that the k nearest neighbor sparse representation classifier has a certain ability to classify the facial expression while greatly reducing the running time,and the average recognition rate and the running time compare with the sparse representation classifier both have improvement.Fourthly,we build Shandong University multi-view facial expression database,which contains six basic expressions of ten peoples under seven angles.The seven web cameras collect the data of seven angles at the same time.In addition,this paper describes several databases which are commonly used in the facial expression recognition,including the frontal face expression database CK + and JAFFE,and the multi-view facial expression database Multi-PIE.And then on the base of these data sets we carry out frontal facial expression recognition experiment and multi-view facial expression recognition experiment.