Research On Face Feature Extraction And Recognition Algorithm Based On Sparse Representation

Face recognition is an important research content in pattern recognition with a very wide scope.Due to various influence factors,such as environment,illumination,expression and posture,face recognition has become a challenging research topic.How to quickly and accurately use computers for face detection and recognition is the key of the current face recognition technology.At present,there are still many problems in face recognition to be solved and improved,which mainly include the completeness of features extraction and the classification performance.Sparse representation based face recognition has attracted much attention due to its simple theory and excellent robustness.In this paper,we mainly study the feature extraction and classification methods based on sparse representation.A large number of experiments show that the proposed algorithms perform well in computational efficiency and recognition rate.The main work and contributions of this dissertation are as follows:(1)A weighted principal component analysis is proposed.The new algorithm weights features by linearly fitting the class labels and features firstly,then makes features' weights be zero by adding a sparse constraint,and finally uses principal component analysis to conduct feature extraction.This method realizes feature pre-selection and highlights important features.(2)Two sparse preserving projections are proposed.One is a weighted sparse neighborhood preserving projections.The new algorithm uses a weighted sparse representation model to learn reconstruction coefficients with restricting the number of nonzero coefficients.This algorithm improves the global robustness.The other is a clustering-based unsupervised discriminant weighted sparse preserving projections.This new algorithm firstly uses clustering to obtain the labels of all training samples and embeds the label information into a discriminant weighted sparse representation.In this way,the recognition accuracy is improved while the simplicity is raised.(3)A discriminant collaborative preserving projections is proposed.This paper proposes a classifier-fitted feature extraction method.The algorithm uses discriminative collaborative representation to construct within-class and between-class graph,which avoids the parameter optimization difficulty,enhances the robustness and improves the recognition performance.(4)A weighted sparse representation classifier is proposed.Original sparse representation based classification classifies the test sample by reconstruction error,which ignore the difference between samples.In this paper,the reconstruction error of a pre-defined reconstruction model is firstly utilized to weight each training samples,and then to solve the weighted sparse representation model.