Research On Facial Expression Recognition Algorithms Based On Sparse Subspace Analysis

Facial expression recognition(FER)is to use computer to automatically complete the classification of the expression,to provide an important basis for the research of human affective computing.Facial expression feature extraction is very omportant for FER,namely how to extract effective features to improve the recognition rates.Linear feature extraction methods are very common in the field of feature extraction.It transforms the original data to a low-dimensional subspace by a linear transformation.And in this subspace,we aim to find the intrinsic low-dimensional features according to certain criterion.However,these obtained features are hard to explain.This paper introduces sparsity to the subspace feature extraction algorithms,to select the effective features automatically,not only can get a good recognition performance,but also can give the reasonal semantic explanation.The main contributions of this paper are listed as follows:1.From the objective function of graph embedding framework,and then based on this,we propose a kind of sparse subspace feature extraction framework,which uses the theory of Fukunage-koontz transform to analyze the distribution of the discriminant information in each subspace,and concluded that discriminant information exists only in the rank space of the total divergence matrix,namely the null space of the total divergence matrix does not contain any discriminant information.The null space can be removed.So we can find the solution in the rank space of the total divergence matrix,and the solution is represented as a matrix equation.By l1-norm restriction,the sparse solution can be obtained by linear Bregman iteration algorithm,at the same time,the convergence of the algorithm is guarantee.Within this framework,we take specific algorithm for example for verifying the validity of the sparse projections.2.On the graph embedding framework,how to construct effective graph is crucial.Inspired by the theory of sparse representation,we use sparse representation model to construct a supervised graph,and propose the double sparse discriminant projection algorithm(DSDP).It mainly utilizes the sparse representation model within each class to construct the graph,the advantage is to estamite the local structure inside the class and to deal with multimodal cases.Finally we uses Fisher criterion to determine the objective function,thus establish the relationship of the algorithm and sparse subspace feature extraction framework stated above.Experiments show that DSDP can obtain better performance than sparse preserving projection(SPP).3.The low rank representation can be regarded as the generalization of sparse representation on the matrix.It can deal with the corrupted images and obtain the clean images.The low rank coefficient matrix in low rank model not only can be seen as the new representation of the original data,but also can be regarded as the similarity among the data points.Thus,we use regularized low-rank model to construct the graph,and propose low rank-based saprse discriminant preserving projection(LRG-SDPP),the algorithm can be thrown into sparse subspace feature extraction framework to obtain the sparse solution.Experiments show that LRG-SDPP is capable of having a good performance for facial expression recognition;especially is robustness to the facial expression recognition with occusions.4.Due to natural representation of the images is tensor form,we propose sparse tensor graph preserving discriminant projection(STGPDP)on the basis of tensor subspace feature extraction.We construct the graph on the tensor data space via the sparse representation,instead of the traditional neighborhood graph based on Euclidean distance.In order to obtain the sparse solution,we convert the objective function into a regression-type problem and then add a regularization or elastic to obtain sparse solution.In the classification phase,the extreme learning machine and nearest neighbor classifier are used to classification and comparison.Experiments show that STGPDP algorithm can achieve better performance than the traditional tensor subspace feature extraction algorithms.Furthermore,the extreme learning machine has more classification ability and faster learning speed than the nearest neighbor classifier.