3D Facial Expression Recognition Based On Facial Features

Facial expression recognition impacts important applications in many areas such as human-computer interactions and data-driven animation.3D face data is robust to pose,illumination,and gestures.Also,it contains more topological characteristic and geometric information than traditional 2D face data,thus can describe the facial muscle movements caused by expressions.In order to improve 3D facial expression recognition performance of featurebased methods,two kinds of recognition algorithms and feature evaluation methods are proposed in this paper.The main work and innovations are as follows:1.A 3D facial expression recognition algorithm using weighted Local Curl Patterns(LCP)is proposed.Firstly,to represent the facial surface changed by expressions,curl vectors are extracted as highly discriminative features based on 3D face.For higher discrimination power of local surface,LCPs are then constructed based on curl vectors.The principle of coding is the same as that of local binary patterns in two-dimension images.Secondly,in order to compute weighs of LCPs,ICNP algorithm is combined with Minimum Projection Error algorithm.Finally,weighted local curl patterns features are entered into the classifier,and then,the expression class of 3D face is predicted.The effect of the algorithm is verified by recognizing 6 different expressions inn the BU-3DFE database.2.A novel 3D facial expression recognition algorithm using SSF-IL-CNN network is proposed.Firstly,convolution kernels are separated into structure parameters and strength parameters.These two parameters are used into initialization and updating respectively.Then,SSF-IL-CNN adopts Island Loss function for conducting strength parameters,which improves the discrimination power of features remarkably.Also,both of depth images and RGB images of 3D face are used for SSF-IL-CNN.The features of them are concentrated as a whole feature for expression recognition.The experiments of the proposed algorithm are carried out on the BU-3DFE datasets,and the results demonstrate that SSF-IL-CNN achieve superior recognition performance.3.Two evaluation methods for features in 3D facial expression recognition are proposed.Firstly,Discrimination Power(DP)is proposed for feature evaluation.And then,a method based on variance and relative entropy is raised.Variance is effective for performing stability of intra-class features,while relative entropy is useful for evaluating difference of inter-class features.However,considering computation costs,this method is not proper for high dimension features,such as convolution features.Therefore,another evaluation method is proposed based on cosine distance,which can also achieve efficient evaluation.The ratio of cosine distance between intra-class features and that between inter-class features is DP.