【作者】 白万荣；

【导师】 王燕；

【作者基本信息】 兰州理工大学 ， 计算机应用技术， 2012， 硕士

【摘要】 人脸识别以其自然、直接、非接触、安全等优点发展为最具潜力的生物特征识别技术，它利用人脸面部特征中的有效信息进行个人身份识别。由于人脸识别在身份验证和识别场合具有巨大的应用价值，以及能促进模式识别等多门学科的发展。因此，对人脸识别技术的研究具有重大的理论和实际意义。提取有效的鉴别特征是人脸识别的一个关键因素，它要求在保持人脸数据集原有的本质结构特性不变的同时进行数据维数约减。流形学习作为一种非线性的维数约减方法，能够有效地学习出高度非线性、属性强相关的高维流形数据的内在几何结构。本文在对基于流形学习的特征提取方法进行深入研究的基础上，主要做了以下工作：1.在邻域保持判别嵌入的基础上，将核映射的思想进行引入其中，并在特征值求解时以Schur正交方式找出最优投影向量，提出了核正交邻域保持判别嵌入算法，克服了邻域保持判别嵌入难以提取非线性特征的困难，很好地保持了人脸流形的几何结构和判别结构信息。2.监督算法和无监督算法都不能充分利用有限的训练样本。因此，本文将无监督判别分析和边界Fisher分析进行结合，改进为半监督算法。其中，利用无监督判别分析来对大量无标签样本进行学习，而利用边界Fisher分析对少量有标签样本进行学习。同时，采用最大散度差准则作为目标函数，避免了散度矩阵奇异值的产生，通过理论分析和实验验证了该方法的可行性和有效性。3.张量边界Fisher分析直接采用图像进行维数约减，避免了传统的方法将图像展开为一维向量的形式，更有效地保持了人脸结构信息。然而在构建最近邻图时，张量边界Fisher分析采用全局统一的k邻域法来选择近邻点的，对于非均匀流形的处理比较困难。本文在研究以上算法的基础上，采用测地距离与欧氏距离的关系来动态的选择训练样本近邻点，使得更有效地选取适合每个样本的局部线性或近似线性区域。更多还原

【Abstract】 Face recognition, characterizing by its naturalness, directness, non-contact,securit, etc., has developed to be a most potential biometric identification technology.The effective information of the facial features is utilized for personal identification.Face recognition has tremendous using value in authentication and identify occasions,and can promote the development of pattern recognition and many other subjects.Therefore, the study of face recognition technology is great theoretical and practicalsignificance.Extract effective discriminant features is a key factor in face recognition, whichrequires reduction of the data dimensionat and keep the face data set of the originalnature of structural characteristics unchanging at the same time. As non-lineardimension reduction methods, manifold learning can effectively learn the intrinsicgeometry of the high-dimensional manifold data structure closely related to highnonlinearity and properties. In this paper, based on the deeply inveatigation of featureextraction method based on manifold learning, the main contents and innovations arelisted as follows:Firstly, on the basis of Neighborhood Preserving Discriminant Embedded (NPDE),introducing the idea of Kernel mapping, finding the optimal projection vector usingSchur orthogonal way while solving the eigenvalue, proposing the Kernel OrthogonalNeighborhood Preserving Discriminant Embedding (KONPDE), overcoming theproblem that NPDE is hard to extract nonlinear characteristics, well maintaining theinformation of geometry and the discriminant of structural information of the facemanifold.Secondly, supervision algorithm and unsupervised algorithm can not make fulluse of limited training samples. Therefore, the Unsupervised Discrimination Projection(UDP) and Marginal Fisher Analysis (MFA) are combined to improve to besemi-supervised algorithm. There among, a large number of label-free samples arestudied using UDP, a small number of label samples are investigated using the MFA.At the same time, choose the maximum scatter difference criterion as the objectivefunction to avoid the divergence matrix singular value, verify the feasibility andeffectiveness of the method through theoretical and experimental alanalysis.Finally, the Tensor Marginal Fisher Analysis (TMFA) use image dimensionalityreduction to avoid that the traditional method expands the image in the form of a one-dimensional vector, which is effectively keep the face structure information.However, when building a nearest neighbor diagram, TMFA employ the global unifiedk-neighborhood method to select a close neighbor of the point, which is more difficultfor non-uniform flow shape processing. On the basis of the above algorithm in theresearch, the method that adopting the relationship between the euclidean distance andthe geodesic distance to select training samples nearest neighbor points dynamically isproposed, which makes it more effective to select local linear or nearly linear regionfor each sample.更多还原