Research On Locality Preserving Subspace Methods For Facial Feature Extraction And Recognition

With the increasing requirements of national defense security and social public security, research on face recognition receives considerable attention. At present, the study of face recognition focuses on the validity of various algorithms. A number of recent research efforts have shown that the face samples possibly reside on a nonlinear submanifold embedding in the high-dimensional image space. On such situation, though they have achieved great success in face recognition, subspace methods such as principal components analysis (PCA) and linear discriminant analysis (LDA), which are effective when the distribution of samples subjects to the Eculidean structure, may lose their validity due to that they are not able to describe the manifold exactly. Compared to PCA and LDA, locality preserving projections (LPP), which is a linear approximation of Laplacian eigenmap, can efficiently describe the manifold of samples and has been widely applied in image indexing and retrieval. However, the classification ability of LPP is weak dues to that it is an unsupervised learning algorithm. In addition, face recognition is a small sample pattern recognition problem; subspace classifier constructed on small sample set is biased and unstable. To these problems, this dissertation investigates the facial feature extraction and recognition methods of locality preserving subspaces by combing subspace analysis methods with classifiers fusion methods, e.g. random subspace method, bagging, etc.The creative works of this dissertation include the following four parts:â‘ The dissertation proposes a random sampling subspace locality preserving projection (RSSLPP) algorithm by combing random subspace method and LPP. RSSLPP algorithm firstly generates multiple random sampling subspaces by randomly sampling the PCA range space under specific strategy and carries out the locality preserving projections in each random sampling subspace. Then, the recognition results in all random sampling subspaces are combined to give the final result. The RSSLPP algorithm efficiently alleviates the drawback that the classification ability of a single unsupervised LPP is weak.â‘¡To extract features that have more discriminative power than LPP features, the dissertation proposes a null space discriminant locality preserving projections (NDLPP) algorithm for the first time. NDLPP algorithm maximizes the modified Fisher criterion and efficiently utilizes the most discriminative information in the null space of locality preserving within-class scatter.â‘¢To address the problem that the discriminant information in the null space of locality preserving within-class scatter decreases with the increasing of training sample number, the dissertation proposes a bagging null space locality preserving discriminant analysis (BagNLPDA) method by combing the bagging algorithm and null space locality preserving discriminant analysis (NLPDA) method. BagNLPDA employs the bootstrap technique of bagging to generate a group of training sample sets, and then the null space locality preserving discriminant analysis is conducted on each traing set. The algorithm not only alleviates the problem that the null space of locality preserving within-class scatter is small when the traing sample number is large but also comes into the characteristic of bagging algorithm that the discriminant information in the null space of locality preserving within-class scatter is utilized multiple times. Thus, the problem that a single NLPDA classifier is biased and has large variance can be alleviated and the classification stability can be improved.â‘£To overcome the problem that both DLPP and NDLPP can not utilize all the discriminative information in the locality preserving subspace, the dissertation proposes a complete discriminant locality preserving projections (CDLPP) algorithm. CDLPP algorithm extracts the regular discriminant features and irregular discriminant features separately in the range space and null space of locality preserving within-class scatter and concatenates them to form the new discriminant features. Thus, CDLPP algorithm utilizes all the discriminant information in the locality presering subspace.Extensive experiments on the standard face databases, e.g. ORL, FERET, Yale, PIE, etc., illustrate that the four locality preserving subspace algorithms proposed in this dissertation are efficient for face recognition. The recognition performances of these four algorithms outperform those of some familiar subspace analysis methods, e.g. Eigenfaces, Fisherfaces, LPP and DLPP. Besides, by the comparison of these four algorithms, it is clear that the classifiers fusion methods, e.g. bagging algorithm and random subspace method, can greatly improve the recognition accuracies of locality preserving subspace methods and make the classification results much stable.