【作者】 张起荣；

【导师】 何中市；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2011， 博士

【摘要】 由于人脸具有唯一性和不易复制的良好特性,并且能在非接触环境和不会对检测人打扰的情况下进行,其商业和实用的价值超过其他的生物特征技术。人脸识别近几十年来越来越受到研究者的重视,使得人脸识别技术发展迅速。在众多人脸识别方法中,子空间方法由于其简单、实施性好、有效性高等特点,逐渐成为近年来的热点研究之一。本文主要对人脸识别特征提取的子空间方法进行研究,并将双树四元数小波应用到人脸识别方向。论文的创新性成果如下:①提出了二维局部鉴别保持投影(2DLDPP)算法。通过对二维局部保持投影(2DLPP)和可调最大间距边缘准则(MMMC)进行研究,发现2DLPP利用了局部结构信息,而MMMC则是利用了全局结构信息。为了使2DLPP方法从无监督的方法变为有监督的方法,本文首次提出了2DLDPP算法。该算法首先是将可调最大间距边缘准则(MMMC)运用到二维中,是二维可调最大间距边缘准则(2DMMMC),然后将2DMMMC引入到二维局部保持投影(2DLPP)中,并且通过构造一个参数来达到最优。2DLDPP是通过2DLPP和2DMMMC进行差的运算,来达到有监督的目的,即保持了2DMMMC的全局结构信息又利用了2DLPP的局部结构信息。所以2DLDPP最大限度地使投影后类内的距离最小化,同时类间的距离最大化。该算法在ORL人脸库和FERET人脸库上进行实验分析,结果表明,2DLDPP方法有效,且优于主成分分析(PCA),二维主成分分析(2DPCA),局部保持投影(LPP)和二维局部保持投影(2DLPP)。②提出了二维双向逆拉普拉斯最大间距边缘准则算法(2DBILMMC)。本文对二维最大间距边缘准则(2DMMC)进行研究,发现2DMMC忽略了局部结构的鉴别信息,而人脸图像是有这部分信息的。为了使2DMMC能够利用上局部结构,本文首次提出2DBILMMC算法。该算法是通过基于图像矩阵的相似距离权重和不相似距离权重来计算出图像的类间拉普拉斯矩阵和图像类内拉普拉斯矩阵,得到二维拉普拉斯最大间距边缘准则,进行逆转,然后合并垂直和水平的信息得到一个复杂的特征矩阵。对二维拉普拉斯最大间距边缘准则进行逆转其实就是让基于特征空间的类内拉普拉斯矩阵与类间拉普拉斯矩阵的最小化,即同类样本距离最小化同时使异类样本最大化。2DBILMMC合并垂直和水平的信息得到一个复杂的特征矩阵,就是充分地利用了两个方向的结构信息。本文首次利用2DBILMMC方法在AR人脸库,FERET人脸库及YALE人脸库上进行实验,结果表明,2DBILMMC有效且优于MMC,2DMMC,LDA和2DLDA。③提出了二维合并双向拉普拉斯逆线性判别分析算法(2DCBLIF)。该算法是通过将拉普拉斯的思想引入到二维的逆线性判别分析算法中得到二维拉普拉斯逆线性判别分析算法,然后合并垂直和水平的信息得到一个复杂的特征矩阵。2DCBLIF合并垂直和水平的信息得到一个复杂的特征矩阵,就是充分利用两个方向的结构信息。本文首次利用2DCBLIF方法在FERET人脸库,CMU PIE人脸库上进行实验,结果表明,2DCBLIF有效且优于2DLDA,2DPCA,2DMMC。④提出了将双树四元数小波(QWT)用于人脸识别方向。QWT是建立在二维希尔伯特变换基础上的。每个四元小波都有一个实部,即离散小波变换和使用四元数代数构造的3个虚部。QWT有3个相位角,其中2个相位角能编码图像的局部变化,另一个相位角则可以编码纹理信息。它可以作为一种新型的分析几何图像特征的多尺度分析工具。本文首次利用QWT在人脸识别。通过在AR和FERET人脸库上的实验结果表明,用QWT处理后的图片再用2DPCA和2DLPP的识别率优于图片直接用PCA,2DPCA,LPP和2DLPP方法。这是一个可能性,即QWT可以分析文本信息丰富的人脸图像。更多还原

【Abstract】 Face recognition has more commercial and practical values than other biometric technologies, due to contactless data acquisition and without any physical damage to the users. Because of this facts face recognition has got more attention, becomes hot research issues, and gains a great development.The subspace analysis-based algorithm is the leading and wide concerns among the present various face recognition algorithms. Because of their’s favorable properties, for example simple, efficient, convenient computation and effectiveness for identification. The subspace analysis-based algorithm has gradually become one of the main methods of facial feature extraction and face recognition. This thesis has mainly focused on the subspace-based feature extraction and dual-tree quaternion wavelets for face recognition algorithms. The main research work and innovations are:A first novel approachcalled two-dimensional locality discriminant preserving projections (2DLDPP) use to the local structure information, and modified maximum margin criterion (MMMC) use to the global structure information are proposed for face feature extraction. This two-dimensional locality preserving projections (2DLPP) and modified maximum margin criterion (MMMC). 2DLPP uses of the local structure information and MMMC. Therefore, in this thesis we introduce the new approach of the 2DLPP method from unsupervised method to a supervised method, therefore, this paper face recognition for image feature extraction as: first, this method is modified maximum margin criterion (MMMC) applied to the two-dimensional, which is two-dimensional modified maximum margin criterion (2DMMMC). Then it is introduced to the two-dimensional locality preserving projections (2DLPP). Finally, it can construct a parameter to achieve the optimum projections. 2DLDPP is conducted by 2DLPP and 2DMMMC with poor operation for the purpose of supervision. It can maintain a 2DMMMC global structural information and use of the 2DLPP local structure information. Extensive experiments are performed on ORL face database and FERET face database. The 2DLDPP method is effective and better than PCA, 2DPCA, LPP and 2DLPP, which can improve performances of face recognition .A second novel approach called two-dimensional bidirctional inverse laplace maximum margin criterion algorithm (2DBILMMC) is proposed for face feature extraction. We found two-dimensional maximum margin criterion (2DMMC) which ignoring the local structure identification of the local structure information. Therefore, for the first time on this thesis we proposed the advantage of the local structure of 2DMMC for this face recognition approach of image feature extraction. The 2DBILMMC can obtain the image within-class Laplacian matrix and image between-class laplacian matrix by the different weighted summation of distances based on image matrices and inverse, then combine vertical and horizontal information into a complex feature matrix. The reversal of two-dimensional laplace maximum margin criterion algorithm is actually minimum the image within-class Laplacian matrix and image between-class laplacian matrix, that is similar samples minimum and dissimilar sample maximum. 2DBILMMC combined vertical and horizontal information to a complex characteristic matrix which is full use of structural information in both directions. Extensive experiments are performed on AR face database, FERET face database and YALE face database. The 2DBILMMC method is effective and better than MMC, 2DMMC, LDA and 2DLDA, which can improve performance of face recognition .A third novel approach called two-dimensional combined bidirctional laplacian inverse fisher discriminate analysis method (2DCBLIF) is proposed for face feature extraction. 2DCBLIF introduces laplacian into inverse fisher discriminate analysis method, then combine vertical and horizontal information into a complex feature matrix.2DCBLIF combined vertical and horizontal information to a complex characteristic matrix is full use of structural information in both directions. Extensive experiments are performed on FERET face database and CMU PIE face database. The 2DCBLIF method is effective and better than 2DLDA,2DPCA,2DMMC,which can improve performance of face recognitionA fourth novel approach dual-tree quaternion wavelets (QWT) is introduced for face feature extraction. QWT is the basis of the 2-D HT, 2-D analytic signal and uses the quaternion algebra. QWT has three phase angles. Two phases can encode local image shifts and the other encodes textural information. Therefore, this paper for the first time thesis introduced this face recognition approach for face feature extraction. Extensive experiments are performed on FERET face database and AR face database. The QWT method achieves better face recognition performance than PCA, 2DPCA, LPP and 2DLPP.更多还原