多流形的人脸特征提取与识别研究

发布时间：2018-03-17 02:02

本文选题：人脸识别　切入点：特征提取　出处：《南昌航空大学》2016年硕士论文　论文类型：学位论文

【摘要】：人脸识别在模式识别与计算机视觉领域中颇受科研人员的热爱,属于生物识别的研究范畴。其中,特征提取是模式识别众多问题中最为重要的一环,人脸识别技术研究的关键所在就是如何提取有利于分类的鉴别特征。传统的全局特征提取方法无法提取人脸图像的局部特征,传统的局部提取方法无法顾及人脸图像的全局特征,并且存在数据维度过高、样本数少和识别效果不理想等问题。本文就基于多流形的特征提取理论和方法做了以下相关研究,主要工作分为以下几部分:(1)阐述了人脸识别的概述、研究背景及内容、应用及难点等,并简要介绍了几种典型的人脸数据库。(2)详细介绍了四种经典的人脸识别特征提取方法:主成分分析(PCA)、线性鉴别分析(LDA)、局部保持投影(LPP)和局部线性嵌入(LLE)方法。并对这四种方法的优缺点进行了简要阐述。(3)在最大间距准则(MMC)算法的基础上,通过引入多流形思想,提出了基于多流形的最大间距准则局部图嵌入(MLGE/MMC)算法。此算法首先构造出多流形外部散度,其次通过多流形内部重建权重矩阵构造出多流形内部散度,最后要达到的目的就是使流形外部的间隔可分性最大以及流形内部的变化最小。与此同时,最大限度地扩大流形边缘,以此更有效地进行特征的提取与分类。此算法采用MMC准则的形式构造目标函数,有效解决了因训练样本较少而导致算法的判别能力下降的问题。(4)非监督线性差分投影(ULDP)方法能使相距比较远的数据点之间的非局部散度达到最大。但ULDP方法也存在着以下不足:1)在学习过程中过分依赖训练样本的数目,当遇到小样本问题时,就严重限制了此方法的应用;2)在提取的众多特征中,无法揭示了哪些特征对分类与预测起到主导作用。为此,我们提出了基于多流形的非监督线性差分投影(MULDP)算法。此算法能得到嵌入在高维空间的低维流形,实现了局部与全局结构信息的有效保持。(5)最后运用MATLAB平台创建了人脸识别系统,来验证经典方法和本文所提出的方法。
[Abstract]:Face recognition is loved by researchers in the field of pattern recognition and computer vision, and it belongs to the research field of biometrics, in which feature extraction is the most important part of pattern recognition. The key of face recognition research is how to extract the discriminant features which are favorable to classification. The traditional global feature extraction method can not extract the local features of face image. The traditional local extraction method can not take into account the global features of face image, and there are many problems such as too high data dimension, fewer samples and less recognition effect. In this paper, the theory and method of feature extraction based on multi-manifold are studied as follows. The main work is divided into the following parts: 1) the overview of face recognition, research background and content, application and difficulties, etc. This paper briefly introduces several typical face database. It introduces in detail four classical face recognition feature extraction methods: principal component analysis (PCA), linear discriminant analysis (LDAA), local preserving projection (LPP) and local linear embedding (LLEs). In this paper, the advantages and disadvantages of these four methods are briefly described. Based on the maximum spacing criterion (MMC) algorithm, this paper gives a brief description of the advantages and disadvantages of the four methods. By introducing the idea of multi-manifold, a local graph embedding algorithm based on the maximum distance criterion of multi-manifold (MLGE / MMC) is proposed. Firstly, the external divergence of multi-manifold is constructed, and then the internal divergence of multi-manifold is constructed by reconstructing the weight matrix of multi-manifold. Finally, the goal to be achieved is to maximize the outer separability of the manifold and minimize the internal variation of the manifold. At the same time, the edge of the manifold is maximized. The algorithm uses the form of MMC criterion to construct the objective function. It effectively solves the problem that the discriminant ability of the algorithm is reduced due to the small number of training samples. (4) the unsupervised linear differential projection (ULDP-) method can maximize the nonlocal divergence between data points far away from each other, but the ULDP method also exists. In the following less than 1: 1) excessive reliance on the number of training samples in the learning process, When the problem of small samples is encountered, the application of this method is severely restricted. Among the many features extracted, it is impossible to reveal which features play a leading role in classification and prediction. We propose an unsupervised linear differential projection (MULDP) algorithm based on multimanifold, which can obtain low dimensional manifold embedded in high dimensional space. Finally, a face recognition system based on MATLAB platform is created to verify the classical method and the method proposed in this paper.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】