Regularized marginal Fisher analysis and sparse representation for face recognition
Huang Ke-kun · Journal of Computer Applications · 2013
When Marginal Fisher Analysis(MFA) is applied to face recognition,it suffers the small size sample problem.If principal component analysis is used to deal with the problem,some useful components will get lost for classification.If replacing the objective function of MFA with maximum margin criterion,it would be difficult to find the optimal parameter.Therefore,in this paper,the regularized MFA method was proposed.It constructed a regularized item by a small number multiplying the identity matrix,and the regularized item was added to within-class scatter matrix so that the resulting matrix was not singular.This method does not lose any useful component for classification and is easy to determine the parameter.Because a sample usually can be linearly represented by few neighbors in the same class,the sparse representation classification was used to further improve the recognition accuracy after regularizing MFA.Experiments were carried out FERET and AR database,and results show that the proposed method can significantly improve the recognition accuracy compared with some classic dimensionality reduction methods.