Research On Robust Facial Expression Recognition Based On Sparse Representation Theory

Facial expression recognition(FER)is a research focus in artificial intelligence and pattern recognition,which belongs to the research contents of affective computing.The researches of FER are applicable to controlled environment at present,which are not ready for automatic facial expression recognition under natural condition.The finding that the firing of the neurons with respect to a given input image is typically highly sparse if these neurons are viewed as an overcomplete dictionary of base signal elements at each visual stage provides a physiological basis for sparse representation.Researchers have employed and developed sparse representation in computer vision and pattern recognition with promising results.This thesis focuses on how to develop sparse representation theory in robust facial expression recognition.The innovative work if this thesis includes:1.From the point view of enhancing sparsity of Sparse Representation-based Classification(SRC),we propose a novel SRC algorithm based on lp-norm(0<p<1)minimization(lp-SRC).Furthermore,we propose an iterative reweighted algorithm to solve the optimization problem of lp-norm minimization.Compared with l1-norm minimization,lp-norm minimization can find a sparser and more accurate solution.Compared with l0-norm minimization,lp-noryr minimization is easier to be solved.Both the theoretical analysis and the experimental results show lp-norm minimization can find a sparser solution.Furthermore,with the decreasing of the value of p.the sparser the solution is.Under the conditions of sufficient training samples or lower feature dimension,the sparser solution is also the more accurate one.Therefore,lp-SRC can achieve better results for robust facial expression recognition.2.By utilizing the similarity between the test sample and all the training samples,we propose a novel algorithm called Similarity Weighted Sparse Representation based Classification(SWSRC).Moreover,a new active-set and feature sign search algorithm is developed to efficiently solve SWSRC.SWSRC can find the optimal sparse coefficients of the test sample over the entire training samples by solving a weighted l1-norm minimization.Compared with the non-convex lp-SRC,SWSRC can find the global optimal solution without iterations.The sparse coefficients of SWSRC solution contains more discriminant information by utilizing the similarity between the test sample and the training samples.So,SWSRC is more effective for facial expression recognition.3.By utilizing the class information,we firstly propose Supervised Canonical Correlation Analysis(SCCA).Secondly,we propose Spectral Supervised Canonical Correlation Analysis(SSCCA)by combing the spectral graph theory with CCA.In SSCCA,we construct an affinity matrix,which incorporates both the class information and local structure information of the data points,as the supervised matrix.SSCCA utilizes the spectral of covariance matrices to obtain a new combined feature,which means it can not only extract the effective information of each single feature,but also eliminate the redundant information within the features,so SSCCA is superior to single feature based method.Further analysis shows that the feature extracted by SSCCA has discriminative information,which implies that SSCCA is suitable for nonlinear recognition problems.We adopt SWSRC to compute the sparse coefficients of the test sample over all training samples,and the classification is performed in the fusion feature space.Furthermore,a unified framework of CCA is proposed to offer an effective methodology for non-empirical structural comparison of different forms of CCA as well as providing a way to extend the CCA algorithm.4.By incorporating correntropy into traditional SRC,we propose Fusion of Correntropy for Sparse Representation based Classification(FCSRC).Furthermore,we develop a novel active-set and feature sign search based algorithm to solve the optimization problem of FCSRC.By combining the global mean square error(MSE)criterion with the local correntropy criterion,the optimal sparse representation coefficients of FCSRC can reveal the relation between the test sample and the training samples locally and globally.Experimental results show that FCSRC can achieve better classification performance in robust facial expression recognition than SRC,especially when the facial images are corrupted.