Manifold Regularized Discriminant Feature Extraction Algorithms With Application To Face Recognition

For the merits of being natural, directly perceived, safe and convenient,face recognition is a biometric technology with great developable potential.Face recognition is also one of the most active research areas, which isclosely related to many disciplines such as Computer Vision, PatternRecognition, Image Processing, Machine Learning and Cognitive Psychology,etc. The study of face recognition technology has both tremendous applicationvalues and important theory significance.A central issue to a successful approach for face recognition is how toextract discriminative features from the facial images. Many featureextraction methods have been proposed and among them the subspaceanalysis has received extensive attention. In a sense, it is commonly acceptedthat human face is a manifold structure. Face dataset is a nonlinear manifoldformed by some inner variables. Face recognition research based on manifoldlearning is attracting more and more attention. Under different algorithmdevelopment frameworks, our research focuses on developing manifoldregularized locality preservation feature extraction algorithms for facerecognition, which can alleviate the out-of-sample, adaptive parameterselection, discriminant extension, nonlinear kernel extension, nonnegativeextension and small sample size (SSS) problems. Each algorithm’seffectiveness has been verified through the simulation experiments onbenchmark face databases.The main contents and originalities of this dissertation are summarized asfollows:1. To overcome the drawback of ignoring local geometric structures ofdata set in LDA and therefore not being able to cope with face recognitionproblems that take the aligned and cropped images with differentillumination and expression variations as input, two locality preserving discriminant analysis methods are developed under the graph embeddingframework. To avoid the SSS problem, a preprocessing dimensionalityreduction method which can reduce the noise effect and enhance theintrinsic difference is adopted in Null Space Discriminative ProjectionEmbedding (NDPE) algorithm and NDPE can seek more discriminativeinformation in the null space of locality preservation within-class scattermatrix. To avoid the eigen-decomposition of dense matrices and reducethe loss of discriminative information, spectral regression technologybased orthogonal locality preserving discriminant mapping algorithm ispresented and we name this algorithm SROLPDM. To improverecognition performance, face recognition methods fusing selectedDTCWT features and original face features at score level or usingcomplex number approach are also developed.2. In manifold learning methods, there exist the out-of-sample problemand the label information of training samples is ignored. Therefore, theoriginal manifold learning methods are not suitable for face recognitionproblems with great illumination, expression or pose variations. Based onthe within-class and between-class local geometry preservation with theirglobal alignment, a novel feature extraction algorithm namedDiscriminant Improved Local Tangent Space Alignment (DILTSA) isproposed in the patch alignment framework. Scatter difference objectivecriterion is adopted in our method and small sample size problem isavoided naturally. To improve the recognition performance, facerecognition methods using augmented Gabor-like wavelet features or thefusion of multi-resolution wavelet features and original face features arealso developed.3. To the adaptive parameter selection and discriminative nonnegativeextension of locality preserving subspace feature extraction for facerecognition with big occlusion or expression variations, based onintegration of modified maximum margin criterion and sparse learning, bydesigning new locality preserving discriminant objective function, a novel feature extraction method named Maximum Margin SparseRepresentation Discriminant Mapping (MSRDM) is proposed. By takingthe part indication virtue of NMF into account, a low-rank approximationalgorithm named Discriminant Neighborhood Margin Regularized NMF(DNMRNMF) is also proposed. DNMRNMF can minimize theapproximation error while contracting intra-class neighborhoods andexpanding inter-class neighborhoods.4. To overcome the drawback of not being able to cope with complex facerecognition problems in linear locality preservation subspace featureextraction methods, a novel feature extraction method named RegularizedKernel Discriminant Local Spline Embedding (RKDLSE) is developed.RKDLSE is a linear approximation and discriminant kernel extensionalgorithm of Local Spline Embedding (LSE). Different methods areadopted to model the within-class and between-class scatter matrices.Regularization method is used to resolve the within-class scatter matrixsingularity. Moreover, with the novel WLTSA nonlinear dimensionalityreduction method and using clustering centers as input, RKDLSE isextended to image set face recognition application.