【作者】 金一；

【导师】 阮秋琦；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2009， 博士

【摘要】 人脸识别是生物特征识别中的一个十分重要的课题,它涉及到图像处理、模式识别、计算机视觉、统计学习、认知学及心理学等众多学科,它也是国家安全和公共安全十分需要和必要的热点问题之一。人脸识别核心算法的研究也因其极高的学术价值和广泛的应用前景,而成为生物特征识别中极富魅力与挑战的课题之一,多年来备受研究人员的关注。有效的提取人脸特征是人脸识别的核心技术之一。能否提取出人脸个体区别于大众的关键特征是高效的人脸识别算法是否能够顺利实现的前提和关键。本文对人脸识别中的若干算法进行了深入研究和探讨,在基于人脸图像表观特征及统计学习理论的模式下,提出了几种人脸识别中的新算法。本文的主要研究内容及创新性工作如下:1.在人脸识别的局部保持投影算法(LPP)的基础上,提出了两种增强局部保持性能的人脸识别新算法1)提出了一种双向压缩变换下的有监督局部保留投影算法。传统LPP需要将二维人脸图像表示成一个较长的一维矢量形式,图像矢量空间的维数过高,使人脸图像特征抽取困难,并容易导致运算复杂及出现奇异矩阵。本文算法在使用2D~2 PCA算法去除图像矩阵行、列相关性后,在有监督模式下最大程度上的保留人脸图像整体信息的同时,直接提取人脸局部邻域结构特征。实验结果表明它降低了计算复杂度与最终表征图像的特征维数,进一步提高了人脸结构特征提取的速度和识别的准确程度;2)提出了一种核正交局部保持投影(KOLPP)的人脸识别算法。KOLPP算法利用核方法提取人脸图像中的非线性信息,并将其投影在一个高维非线性空间,在保证各向量正交的同时,通过局部保持投影算法做线性映射,从而有效地提取人脸非线性局部邻域结构特征。该算法采用有监督模式增强了人脸局部保持性能,同时,正交化约束条件的加入最大程度上的提取人脸之间的特性,因而能够很好地发掘人脸图像中的高维非线性结构,获得较为理想的识别结果。2.针对LSDA算法的优缺点进行深入的分析与研究,提出了正交局部敏感判别分析(OLSDA)和张量正交局部敏感判别分析(Tensor-OLSDA)1)OLSDA算法首先将人脸数据映射在一个线性子空间中,该子空间通过最大化数据点属于同一类的近邻点与属于不同类的近邻点之间在局部邻域结构上的差值边缘图,而达到同时保持人脸局部结构及判别式信息。随后,正交基函数作为附加约束直接作用于目标函数,增强了不同类间的判别信息。实验结果也证明了算法的有效性和稳定性;2)在OLSDA的基础上,提出了张量正交局部敏感判别分析,该算法将人脸图像表示表示成高阶Tensor的形式,更有利于保持人脸图像作为一个二维矢量的空间信息,并且基于Tensor的表示形式不需要将数据展开成一个高维矢量,有效的解决了矩阵奇异性的问题,实验结果也显示了该算法能进一步提高人脸识别的准确率,获得较理想的识别结果。3.提出了一种旋转与平移不变的联合子空间人脸识别算法(Rotate and ShiftInvariant-based United Subspace Analysis,Rotate and Shift Invariant-based USA)对姿态与距离变化的人脸识别,提出了一种快速、有效的,针对旋转与平移不变的联合子空间新方法。该算法将局部特征及细节纹理信息增强了的人脸Gabor特征通过双方向的二维主成分算法(2D~2 PCA)进行整体特征提取及降维处理,最后进一步使用核心算法一:监督局部保持投影算法(United-SLPP),及核心算法二:正交局部敏感判别式分析算法(United-OLSDA)进行二次特征提取,并且在两个不同规模的人脸库上对这种联合子空间算法在正确识别率及识别速度进行了测试分析。实验结果显示,本论文提出的旋转与平移不变的联合子空间人脸识别算法,结合了前文所提算法的综合优势,在不同条件下,均能明显提高人脸识别的准确程度。更多还原

【Abstract】 Face recognition is a key subject of the research on biometrics.It relates to image processing,pattern recognition,computer vision,statistical learning,cognitive science and psychology,and many other important disciplines.It is also in great need and necessary of national security and public safety,and becomes one of the hottest issues in that field.Due to great academic research value and prospects of a wide range of applications,the research of effective and efficient face recognition algorithms has become one of the most attractive and challenging tasks in biometrics,and has gained wide attentions by researchers both at home and abroad for years.Effective feature extraction is one of the core technologies for face recognition.To extract the most distinctive face features of the same individual,which is different from the general public,is the first and crucial phase for a highly efficient face recognition algorithm.In this dissertation,facial feature extraction and pattern classification for the human faces have been studied and discussed.Aiming at the difficulty of the traditional appearance-based pose estimation,several new approaches are proposed under the frame of statistic learning.The main research content and innovative work are as follows:1.Two kinds of face recognition algorithms which aim at enhancing the locality preserving performance are proposed based on the research of Locality Preserving Projections(LPP).1) A Supervised locality preserving projections under Bi-directional Compression Transformation(SLPP-BCT) algorithm is proposed for face recognition.The traditional LPP represents a face image by a vector in high-dimensional space which leads to the difficulty of feature extraction and is easily confronted with the matrix singular problem and high computational complexity.In this new proposed method,the bilateral-projection-based 2DPCA(2D~2 PCA ) algorithm is used to remove the redundancy from two directions of the image.It preserves the face image structure as a whole,meanwhile,it directly preserve the local information of the compressed data space under a surprised mode. Experiments demonstrate the effectiveness and efficiency of the new proposed method.It outperforms some most popular algorithms on both recognition speed and accuracy.2) A new method called kernel based orthogonal locality preserving projections algorithm is proposed for face representation and recognition.In this method, the nonlinear kernel mapping is used to map the face data into an implicit feature space,and then a linear transformation which produces orthogonal basis functions is performed to preserve locality geometric structures of the face image.KOLPP is performed under a supervised learning mode which improved the locality preserving capacity of the face samples,and the orthogonalization constraints enhanced the discriminated features extraction between different individuals,simultaneously.Therefore,KOLPP algorithm preserves the nonlinear geometric structures of face image better and obtains a more satisfactory recognition performance.2.Two novel appearance-based methods,called Orthogonal Locality Sensitive Discriminant Analysis and Tensor-based Orthogonal Locality Sensitive Discriminant Analysis(Tensor OLSDA),are proposed based on the analysis of the newly proposed Locality Sensitive Discriminant Analysis(LSDA) algorithm.1) OLSDA projects the face data into a linear subspace which maximized the margin constructed by data points from the same class and the different classes at each local neighborhood,so that preserves not only the local neighborhood information but discriminant information as well.Furthermore,the orthogonal basis function based constraint is added into the objective function of LSDA to emphasize the discriminant information.Orthogonal LSDA algorithm is proposed to preserve the local geometrical structure by computing the mutually orthogonal basis functions iteratively.Experimental results also proved its validity and stability.2) Motivated by the Locality Sensitive Discriminant Analysis(LSDA),a novel appearance-based method that called Tensor Orthogonal Locality Sensitive Discriminant Analysis(Tensor OLSDA) is presented for face recognition.With face data’s high-order tensor representation,this new method preserves its spatial structure of the face image better,which is actually in a 2-D vector form. Tensor-based representation doesn’t need to expand the face data into a high-dimensional space which avoids the problem of singular matrix effectively.Experimental results also show the impressive performance of the proposed method.3.The Rotate and shift invariant based United Subspace Analysis(Rotate and Shift Invariant-based USA) is proposed in this thesis.A fast and effective new method,called Rotate and Shift Invariant based United Subspace Analysis,is proposed for the pose and distance changed face recognition in this paper.In the proposed method,the local characteristics and detail texture information enhanced Gabor feature is first compressed by the 2D~2 PCA algorithm to extract the global feature and reduce dimension of the face features.Then the core algorithmⅠ(United-SLPP) or the core algorithmⅡ(United-OLSDA) are utilized for further feature extraction,respectively.Moreover,the united subspace algorithm is tested and analyzed with the experiments on two face databases of different scales.The results show that,the proposed Rotate and Shift Invariant based United Subspace Analysis takes the comprehensive advantages of the algorithms proposed above and can significantly improve the accuracy of face recognition in different occasions.更多还原