Research On 3D Facial Expression Recognition Algorithm Based On Convolutional Neural Network

Facial expression is a means to express human cognition,emotion and state.Accurate and effective facial expression recognition plays an important role in promoting natural and harmonious human-computer interaction.Therefore,automatic facial expression recognition based on machine has a wide range of applications in human-computer interaction,facial animation and psychological research.At present,2D facial expression recognition methods that rely on facial texture analysis are still largely affected by illumination and attitude changes.In recent years,with the rapid innovation of 3D data acquisition equipment,3D technology is considered as a promising method for robust facial expression analysis.The 3D data is not sensitive to light and is not affected by attitude changes,which is needed for more robust facial expression recognition processing.This thesis studies how to use convolutional neural network to realize efficient classification of 3d facial expressions.The main research work is as follows:1.A new deep feature fusion convolutional neural network(CNN)model is proposed to learn the combination relation and weight of different features of 3d face.Firstly,the normal vector and curvature of 3d face point cloud are calculated.Then,with the combination of different face attribute features,the depth feature was adjusted to fuse with CNN subnet to learn 3d face features.The results showed that the recognition rate of depth,normal and texture features was the highest,reaching 79.17%.2.In order to solve the problem of high algorithm complexity in the traditional method,the conformal mapping algorithm is used to transform the 3d model of the face into a 2d feature graph related to the attribute information(the connection relation between color,face and vertex)and take it as the input of the neural network.Finally,the feature visualization is used to compare the expression classification of 2D feature map and 2D texture image obtained by using conformal mapping algorithm,which proves the advantages and effectiveness of the proposed algorithm.3.A new CNN model based on transfer learning(CTL-CNN)is proposed to learn the 2d attributes of 3d face data obtained by conformal mapping.The model learns facial expression features by fine-tuning the pre-trained deep CNN subnet.The experiment uses BU-3DFE database,and the results show that the recognition rate ofthe proposed method is 92.50% in subset I and 89.58% in subset II.