基于图嵌入与弹性网络回归的特征提取算法及其在人脸识别中的应用

发布时间：2018-04-20 18:20

本文选题：人脸识别 + 特征提取　；参考：《南昌航空大学》2017年硕士论文

【摘要】：在人脸识别过程中,特征提取的重点在于挖掘并提取人脸数据中的关键特征,这有利于提高算法的识别和分类能力。传统基于子空间学习的特征提取算法如主成分分析(PCA)和线性判别分析(LDA),以及基于流形学习的图嵌入特征提取算法如局部线性嵌入(LLE)和局部保持投影(LPP),因为具有简单、直观、高效等优点被广泛使用。但是上述算法仍然存在许多问题和局限性,例如不能同时得到数据的全局和局部结构、线性方法对于非线性数据处理不理想、“小样本”问题以及特征冗余等等。基于稀疏特征提取的研究是人脸识别领域中的另一个热点。原始人脸数据中往往包含众多特征,稀疏特征提取可以从原始数据中找到某些最显著的特征,然后使用它们组成最小特征子集对原始数据进行最优表示,这一过程既可以简化数据又能够保留数据中的关键信息。弹性网络回归(Elastic Net)是目前常用的稀疏特征提取算法之一。本文结合常用图嵌入算法和弹性网络回归,针对上述特征提取算法中存在的问题进行研究,提出新的算法并应用在人脸识别中,主要工作有:(1)简单介绍了人脸识别的研究背景及发展历程、研究内容及应用、存在的问题等,并对几种典型的人脸数据库作了简要说明;(2)根据本文研究的内容,分别介绍了基于流形学习的图嵌入以及稀疏特征提取的思想,并对经典特征提取算法(PCA、LDA、LLE和LPP)以及稀疏特征提取算法(岭回归(Ridge)、套索回归(Lasso)和弹性网络回归)的实现步骤进行了细致的介绍,然后简单分析了上述算法的优缺点;(3)结合PCA、LLE以及弹性网络回归,提出了无监督稀疏差分嵌入(USDE)特征提取算法。该算法的基本思想是:首先,构建出基于LLE的“局部最小嵌入”以及基于PCA的“全局最大方差”;然后,使用“差分”形式解决多目标最优化问题,并结合稀疏约束构建USDE目标函数;最后,使用弹性网络回归进行稀疏性实现;(4)在最大边界准则(MMC)算法的基础上,结合LLE和弹性网络回归提出了基于最大边界准则的稀疏局部嵌入(SLE/MMC)算法。首先,SLE/MMC在保持局部近邻的基础上构建类内散布矩阵以及类间散布矩阵;然后,SLE/MMC使用“MMC”的形式以及稀疏约束构造SLE/MMC的目标函数;最后,SLE/MMC使用弹性网络回归得到一个稀疏化的结果。(5)结合二维判别局部保持投影(2DDLPP)和弹性网络回归,提出了基于稀疏二维判别局部保持投影(S2DDLPP)的特征提取方法。2DDLPP在LPP中引入类间离散度和类别信息,并直接利用原始人脸数据矩阵而不是变换后的向量进行特征映射,可以减少在变换过程中的信息损失。首先,在2DDLPP基础上,S2DDLPP在满足“类内距离最小化”和“类间距离最大化”的同时,在其目标函数上加入稀疏约束;然后,S2DDLPP使用弹性网络回归进行稀疏性实现,得到一个最优稀疏投影矩阵。
[Abstract]:In the process of face recognition, the emphasis of feature extraction is to mine and extract the key features from face data, which is helpful to improve the recognition and classification ability of the algorithm. Traditional feature extraction algorithms based on subspace learning, such as principal component analysis (PCA) and linear discriminant analysis (LDAA), and manifold learning based graph embedding feature extraction algorithms, such as local linear embedding (LLEE) and locally preserving projection (LPP), are simple and intuitive. The advantages of high efficiency are widely used. However, these algorithms still have many problems and limitations, such as the global and local structure of the data can not be obtained simultaneously, the linear method is not ideal for dealing with nonlinear data, the "small sample" problem and the feature redundancy and so on. The research of sparse feature extraction is another hotspot in the field of face recognition. Primitive face data often contain many features. Sparse feature extraction can find some of the most prominent features from the original data, and then use them to form a minimal feature subset to represent the original data optimally. This process can both simplify the data and retain the key information in the data. Elastic Networked is one of the commonly used sparse feature extraction algorithms. In this paper, based on the commonly used graph embedding algorithm and elastic network regression, the problems existing in the above feature extraction algorithms are studied, and a new algorithm is proposed and applied to face recognition. The main work is: (1) briefly introduces the research background and development course, research contents and applications, existing problems of face recognition, and gives a brief description of several typical face databases. The ideas of graph embedding and sparse feature extraction based on manifold learning are introduced respectively. The implementation steps of the classical feature extraction algorithms (PCALDALE and LPP) and sparse feature extraction algorithms (Ridgeg Ridge, Lassoand Elastic Network regression) are introduced in detail. Then, the advantages and disadvantages of the above algorithms are briefly analyzed. Combined with PCALLE and elastic network regression, an unsupervised sparse difference embedding (USDE) feature extraction algorithm is proposed. The basic ideas of the algorithm are as follows: firstly, the "local minimum embedding" based on LLE and the "global maximum variance" based on PCA are constructed; then, the "difference" form is used to solve the multi-objective optimization problem. The objective function of USDE is constructed with sparse constraints. Finally, using elastic network regression to implement sparsity, we use the maximum boundary criterion (MMC) algorithm. Combined with LLE and elastic network regression, a sparse local embedding (SLER / MMC) algorithm based on the maximum boundary criterion is proposed. Firstly, the intraclass dispersion matrix and the inter-class dispersion matrix are constructed on the basis of preserving the local nearest neighbor, and then the SLER / MMC uses the form of "MMC" and sparse constraints to construct the objective function of SLE/MMC. Finally, SLER / MMC uses elastic network regression to obtain a sparse result. 5) combining with 2-D discriminant local preserving projection 2DDLPP) and elastic network regression, This paper presents a feature extraction method based on sparse two-dimensional discriminant locally preserving projection (S2DDLPP). 2DDLPP introduces inter-class dispersion and class information into LPP, and directly uses primitive face data matrix instead of transformed vector to map features. It can reduce the loss of information in the process of transformation. Firstly, on the basis of 2DDLPP, S2DDLPP not only satisfies "minimization of intra-class distance" and "maximization of inter-class distance", but also adds sparse constraints to its objective function, then S2DDLPP uses elastic network regression to implement sparsity. An optimal sparse projection matrix is obtained.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】