Face Recognition Based On Manifold Learning And Its Improved Algorithm

Human face recognition is not only a hot research subject in the domain of model recognition and artificial intelligence, but also a most valuable and potential recognition technology by biometric characteristic for the merits of being natural, directly perceived,safe and convenient. The study of face recognition technology has both important theory significance and tremendous application values.A central issue to a successful approach for face recognition is how to extract discriminative features from the facial images. Many feature extraction methods have been proposed and among them the subspace analysis has received extensive attention.In a sense, it is commonly accepted that human face is a manifold structure. Face dataset is a nonlinear manifold formed by some inner variables. Face recognition research based on manifold learning is attracting more and more attention. In this thesis,we introduce some classical manifold learning algorithms firstly. Then, under different algorithm development frameworks, this article focuses on improving and extending the classical manifold learning algorithms, which can alleviate the limits of Euclidean distance in the manifold structure, not considering the face data category information and face data may exist multiple manifold structure problems. The proposed algorithms’ effectiveness has been verified through the simulation experiments on benchmark face databases.The main contents and contributions of this paper can be generalized as follows:① In the algorithm named locality preserving projections(LPP), only using Euclidean distance to calculate the face manifold’s neighbor structure may produce false close neighbor relations. Also LPP uses the way by consolidates partial of data to reconstruct the overall intrinsic properties of face, but have not consider to take measures to increase the distribution of the differences between heterogeneous data. To solve these problems, correlation coefficient fused with LPP algorithm(CCLPP) is proposed, the data category information is introduced by correlation method into the algorithm, which uses correlation coefficient fused with Euclidean distance to build and estimate neighborhood graph. CCLPP algorithm can reveal the true intrinsic local geometric structure and the overall data distribution characteristics of the face datasets.On the basis of maintaining the real neighborhood structure, expanding the spacing between the different types of data. Thus improving the false neighborhood of the facedata samples, the new method shows a good characteristic of identification.② The existing manifold learning methods are all based on single manifold hypothesis, considering all of the data samples distributed in a unitive embedded low dimensional manifold. New research suggests that facial images may present multiple manifold distribution, data from different classes may reside on different manifolds of possibly different dimensionality. If we describe the face data which reside on multiple manifolds in a single manifold will affect the estimation of the true geometric structure of the face data. This paper probed into a general framework for data classification on multiple manifolds and proposed a new algorithm named Multiple Neighborhood Preserving Embedding(M-NPE), which expects face data reside on multiple manifolds by their classes. In the steps of M-NPE algorithm, the intrinsic features are firstly learned for each class separately by NPE algorithm, and the Genetic algorithm is then employed to get the near optimal dimensionality of each manifold from the classification viewpoint. Finally in M-NPE, new data samples can be classified according to the minimum reconstruction error on multi-manifolds.