用于人脸识别的LDA方法的研究及其应用
摘要
随着社会的发展 ,各个方面对快速有效的自动身份验证的要求日益迫切。由于生物特征是人的内在属性,具有很强的自身稳定性和个体差异性,是身份验证的最理想依据。从而，“生物特征识别技术”，因其良好的安全性越来越多地应用于身份识别。自动人脸识别技术因造价低、使用友好等优点成为其中很有前景的一部分。由于在一个场景中找到一张脸并且能够识别它在人类生活中是很普通且非常重要的，因此将这项任务自动化是有重要意义的。计算机技术和人工智能领域研究的发展使得这方面的研究成为可能。同时，根据计算机技术的发展趋势，智能人机界面应该取代传统的键盘鼠标，成为更加安全方便的人机交互接口。
人脸识别是一个跨学科非常具有挑战性的前沿课题。首先因为人脸图像的获取过程不同，导致二维图像信息在质量、几何、光线上都有内在的不同，此外还有脸部受到遮挡和化妆等因素的影响。但是，更内在的原因是，人脸是具有高度相似性的非刚体。人脸不同于普通物体，不同人的脸具有高度的相似性，同一人的脸又具有不同的状态，不仅存在各种变形，而且和皮肤之间是平缓过渡，这使得人脸识别问题不同于普通物体的识别问题。目前，许多研究机构致力于这一领域的研究，取得了丰硕的理论成果并有不同的应用软件应运而生。尽管如此，可以说，还没有一个完全的解决方案可以在毫无约束的情况下出色地完成人脸识别任务。
在应用领域上，目前可能的识别任务主要可以归为两类：身份识别/辨认/匹配(Recognition/ Identification/ match)、身份验证/证实/监督(Verification/ Authentification /Surveillance)。前者在应用上的典型实例就是公安刑侦追逃。第二种应用情形的典型实例是身份证件的鉴别、自动门禁控制系统、银行ATM取款机、家庭安全等领域。简单讲，一个是判断他或她是谁，一个是判别他或她是不是某个人。目前，这一领域的产品正在受到越来越多的关注。
本文在已有的理论研究成果的基础上做了进一步的研究和实验工作，并将这些工作部分转化成软件产品，取得了一定的应用。
一个自动人脸识别系统的主要组成部分是图像获取、人脸的检测定
位、人脸图像预处理、特征提取、分类器设计和决策部分。较为重要和困难的问题集中在人脸检测定位，特征提取，分类器设计上。
本文针对实际应用的要求，对检测定位部分采用多分类器级联结构实现[1]。由积分图像可以快速、容易地计算出大量的简单特征，再用AdaBoost学习算法挑选一些重要的特征，并构造一系列的弱分类器，多个弱分类通过线性组合可构造出一个强分类器。采用一个分类器逐渐复杂的多分类器级联结构大大提高了检测速度，在奔Ⅲ933以上的机器上，处理速度可达每秒20帧以上。利用改进的基于前向序列特征选择的Adaboost算法，采用代回溯过程的FloatBoost的特征选择算法。采用简单的块特征以及一个金字塔的分级结构，该系统可以检测多视角人脸图像，处理速度达每秒5帧以上。
特征提取和选择部分采用统计分析领域的线性判别式分析(LDA)方法。这种方法的基本思想是将测量空间中的高维数据，通过LDA训练过程得到线性变换矩阵，投影到最佳鉴别矢量空间已达到维数压缩的效果，投影后保证模式样本在新的空间中有最大的类间距离和最小的类内距离。即模式在该空间中有最佳的可分离性。这是一种有效的模式分析技术。这种方法的关键是如何求解最佳变换矩阵。这种方法在最少损失判别信息的前提下获得了最大的数据降维，是简单而有效的降维方法。本文中尝试了对这种经典的模式识别算法的一些改进算法。
分类策略选取的是加权的欧式距离。同时也研究比较了近年来流行的多种分类器，有CityBlock方法、欧氏距离、斜方差法、关联规则方法、马氏距离。人脸识别系统中使用的是加权的欧氏距离方法，其他几种方法只作为实验研究。
在上述的理论和实验的基础之上，我们研制了可以用于实际生产生活的软件产品――自动人像屏幕保护系统。此外，还对本项目组从前研究开发的大型图像库检索等产品作了二次开发。这种产品是身份验证类的应用，来访者坐在显示器前，系统利用摄像头拍的照片，通过图像判断此人是否有权限通过验证进入计算机系统。后者是身份识别类的应用，对输入的人像，找出最为匹配的一个图像集。
由于时间的关系，很多理论还没有进行深入的挖掘；人脸表示方面的研究不够充分，没有对比实验；对人像库检索这一类的应用没有作出在更大规模人像库的情况下的解决方案，这些都是值得今后继续研究的问题。
Biometric identification systems, which use physical features to check a person's identity, ensure much greater security than password and number systems. Identifying a human individual from his or her face is one of the most non-intrusive modalities in biometrics. The capability of finding and recognizing a face in a random scene is important in everyday activities. So it’s very significant that automate this task. Development in computer technology and artificial intelligence make research on these fields possible. At the same time, it also spurs the efforts on intelligent interface between human and machine. It is fundamental that computer should know who are in its eyeshot.
However, it is also one of the most challenging problems. At first, the face images are obtained by different way or in different condition, so they have substantial difference in quality, geometry, illumination, etc. In addition, it also exits the makeup and face-painting influence. But the most essential reason is that face is a kind of non-rigid object that has highly similarity. Different person’s faces have similar shape and structure, and one person’s face has different state. In the past decade, many research groups make great efforts on it and a series of successes have made general personal identification appear not only technically feasible but also economically practical. However, no perfect solution can accomplish this task under the non-constraint condition.
It is of particular interest in a wide variety of applications. In fact, face recognition technology has two kinds of application: recognition/ identification/ match and verification/ authentification/ surveillance. A typical example about the former one is applications in law enforcement for mug-shot identification. The latter application is also broad, such as verification for personal identification, gateways to limited access areas, authentification for ATM and family security, etc. In brief, one is concluding who is he/she and the other is deciding is he/she somebody.
This paper is interested in doing some father work in theory research and practical experiment. It also outlined two instances concerned.
A typical automatic face recognition system consists of the following function: image obtaining, face detection and location, image preprocessing, feature extraction, classifier designing and deciding. Most research focuses on face detection, feature extraction and classifier designing.
According the practical application, an integrated, robust,real-time face detection and demographic analysis system. Faces are detected and extracted using the fast algorithm recently proposed by Viola and Jones . Detected faces are passed to a demographics classifier which uses the same architecture as the face detector.
Linear Discriminant Analysis (LDA), or Fisher’s Linear Discriminant (FLD) is a class specific method in the sense that it represents data to make it useful for classification. Given a set of images with each image belongs to one of classes , LDA selects a linear transformation matrix in such a way that the ratio of the between-class scatter and the within-class scatter is maximized.
We select classical Euclidean Distance method for recognition strategy. It is simple and efficient in many applications. However computing in recognition phrase is excessive. In order to achieve excellent performance, each axis is weighted by LDA＇s associated generalized eigenvalue.In- addition,other methods are also studied and tested.
Based on the above theory and experiment, software has been developed to practical application. Automatic face recognition screensaver system is the kind of verification application. The system achieves the visitor’s face image then decides whether he/she has the access permission by matching the images features in database. The other software is concerned recognition application. It’s face image database searching system. Given an image, system finds an image set, which includes the most similar images.
Some possible future research
人脸识别是一个跨学科非常具有挑战性的前沿课题。首先因为人脸图像的获取过程不同，导致二维图像信息在质量、几何、光线上都有内在的不同，此外还有脸部受到遮挡和化妆等因素的影响。但是，更内在的原因是，人脸是具有高度相似性的非刚体。人脸不同于普通物体，不同人的脸具有高度的相似性，同一人的脸又具有不同的状态，不仅存在各种变形，而且和皮肤之间是平缓过渡，这使得人脸识别问题不同于普通物体的识别问题。目前，许多研究机构致力于这一领域的研究，取得了丰硕的理论成果并有不同的应用软件应运而生。尽管如此，可以说，还没有一个完全的解决方案可以在毫无约束的情况下出色地完成人脸识别任务。
在应用领域上，目前可能的识别任务主要可以归为两类：身份识别/辨认/匹配(Recognition/ Identification/ match)、身份验证/证实/监督(Verification/ Authentification /Surveillance)。前者在应用上的典型实例就是公安刑侦追逃。第二种应用情形的典型实例是身份证件的鉴别、自动门禁控制系统、银行ATM取款机、家庭安全等领域。简单讲，一个是判断他或她是谁，一个是判别他或她是不是某个人。目前，这一领域的产品正在受到越来越多的关注。
本文在已有的理论研究成果的基础上做了进一步的研究和实验工作，并将这些工作部分转化成软件产品，取得了一定的应用。
一个自动人脸识别系统的主要组成部分是图像获取、人脸的检测定
位、人脸图像预处理、特征提取、分类器设计和决策部分。较为重要和困难的问题集中在人脸检测定位，特征提取，分类器设计上。
本文针对实际应用的要求，对检测定位部分采用多分类器级联结构实现[1]。由积分图像可以快速、容易地计算出大量的简单特征，再用AdaBoost学习算法挑选一些重要的特征，并构造一系列的弱分类器，多个弱分类通过线性组合可构造出一个强分类器。采用一个分类器逐渐复杂的多分类器级联结构大大提高了检测速度，在奔Ⅲ933以上的机器上，处理速度可达每秒20帧以上。利用改进的基于前向序列特征选择的Adaboost算法，采用代回溯过程的FloatBoost的特征选择算法。采用简单的块特征以及一个金字塔的分级结构，该系统可以检测多视角人脸图像，处理速度达每秒5帧以上。
特征提取和选择部分采用统计分析领域的线性判别式分析(LDA)方法。这种方法的基本思想是将测量空间中的高维数据，通过LDA训练过程得到线性变换矩阵，投影到最佳鉴别矢量空间已达到维数压缩的效果，投影后保证模式样本在新的空间中有最大的类间距离和最小的类内距离。即模式在该空间中有最佳的可分离性。这是一种有效的模式分析技术。这种方法的关键是如何求解最佳变换矩阵。这种方法在最少损失判别信息的前提下获得了最大的数据降维，是简单而有效的降维方法。本文中尝试了对这种经典的模式识别算法的一些改进算法。
分类策略选取的是加权的欧式距离。同时也研究比较了近年来流行的多种分类器，有CityBlock方法、欧氏距离、斜方差法、关联规则方法、马氏距离。人脸识别系统中使用的是加权的欧氏距离方法，其他几种方法只作为实验研究。
在上述的理论和实验的基础之上，我们研制了可以用于实际生产生活的软件产品――自动人像屏幕保护系统。此外，还对本项目组从前研究开发的大型图像库检索等产品作了二次开发。这种产品是身份验证类的应用，来访者坐在显示器前，系统利用摄像头拍的照片，通过图像判断此人是否有权限通过验证进入计算机系统。后者是身份识别类的应用，对输入的人像，找出最为匹配的一个图像集。
由于时间的关系，很多理论还没有进行深入的挖掘；人脸表示方面的研究不够充分，没有对比实验；对人像库检索这一类的应用没有作出在更大规模人像库的情况下的解决方案，这些都是值得今后继续研究的问题。
Biometric identification systems, which use physical features to check a person's identity, ensure much greater security than password and number systems. Identifying a human individual from his or her face is one of the most non-intrusive modalities in biometrics. The capability of finding and recognizing a face in a random scene is important in everyday activities. So it’s very significant that automate this task. Development in computer technology and artificial intelligence make research on these fields possible. At the same time, it also spurs the efforts on intelligent interface between human and machine. It is fundamental that computer should know who are in its eyeshot.
However, it is also one of the most challenging problems. At first, the face images are obtained by different way or in different condition, so they have substantial difference in quality, geometry, illumination, etc. In addition, it also exits the makeup and face-painting influence. But the most essential reason is that face is a kind of non-rigid object that has highly similarity. Different person’s faces have similar shape and structure, and one person’s face has different state. In the past decade, many research groups make great efforts on it and a series of successes have made general personal identification appear not only technically feasible but also economically practical. However, no perfect solution can accomplish this task under the non-constraint condition.
It is of particular interest in a wide variety of applications. In fact, face recognition technology has two kinds of application: recognition/ identification/ match and verification/ authentification/ surveillance. A typical example about the former one is applications in law enforcement for mug-shot identification. The latter application is also broad, such as verification for personal identification, gateways to limited access areas, authentification for ATM and family security, etc. In brief, one is concluding who is he/she and the other is deciding is he/she somebody.
This paper is interested in doing some father work in theory research and practical experiment. It also outlined two instances concerned.
A typical automatic face recognition system consists of the following function: image obtaining, face detection and location, image preprocessing, feature extraction, classifier designing and deciding. Most research focuses on face detection, feature extraction and classifier designing.
According the practical application, an integrated, robust,real-time face detection and demographic analysis system. Faces are detected and extracted using the fast algorithm recently proposed by Viola and Jones . Detected faces are passed to a demographics classifier which uses the same architecture as the face detector.
Linear Discriminant Analysis (LDA), or Fisher’s Linear Discriminant (FLD) is a class specific method in the sense that it represents data to make it useful for classification. Given a set of images with each image belongs to one of classes , LDA selects a linear transformation matrix in such a way that the ratio of the between-class scatter and the within-class scatter is maximized.
We select classical Euclidean Distance method for recognition strategy. It is simple and efficient in many applications. However computing in recognition phrase is excessive. In order to achieve excellent performance, each axis is weighted by LDA＇s associated generalized eigenvalue.In- addition,other methods are also studied and tested.
Based on the above theory and experiment, software has been developed to practical application. Automatic face recognition screensaver system is the kind of verification application. The system achieves the visitor’s face image then decides whether he/she has the access permission by matching the images features in database. The other software is concerned recognition application. It’s face image database searching system. Given an image, system finds an image set, which includes the most similar images.
Some possible future research
引文
[1] Rapid Object Detection using a Boosted Cascade of Simple Features Paul Viola Michael Jones viola@merl.com mjones@crl.dec.com Mitsubishi Electric Research Labs Compaq CRL 201 Broadway, 8th FL One Cambridge Center Cambridge, MA 02139 Cambridge, MA 02142
[2] Kah-Kay Sung, Tomaso Poggio, "Example-based Learning for View-based Human Face Detection" , Massachusetts Institute of Technology, A.I. Memo No.1521 , C.B.C.L Paper No.112,1994
[3] 人脸检测研究进展 刘党辉 沈兰荪 文章编号1002－8331－（2003）28－0005－05 文献标识码 - A 中图分类号 TP391
[4] Viola P,Jones M.Fast and Robust Classification Using Asymmetric Adaboost and a Detector Cascade[R]. http://www.merl.com/people/viola/research/publications/2002.
[5] 边肇祺，张学工等编著，《模式识别》，清华大学出版社，2000年，第二版，p223－p227
[6] Liu K,ChengYQ,Yang JY.An efficient algorithm for Foley-Sammon optimal set of discriminant vectors byalgebraic method. International Journal of Pattern Recognition and Artificial intelligence,1982,6(5): 817～829.
[7] Viola P,Jones M.Robust Real Time Object Detection[C].In :Proc of IEEE Work shop on SCTV,2001.
[8] An Efficient LDA Algorithm for Face Recognition.Jie Yang, Hua Yu, William Kunz School of Computer Science Interactive Systems Laboratories Carnegie Mellon University Pittsburgh, PA 15213
[9] M.Turk andA. Pentland.Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):72–86, 1991.
[10]K.Etemad and T.Chellappa,“Discriminant Analysis for Recognition of Human Face Images,”Journal of the Optical Society of America A,Vol.14,pp.1724-1773,1997.
[11]Kah-Kay Sung, Tomaso Poggio, "Example-based Learning for View-based Human Face Detection" , Massachusetts Institute of Technology, A.I. Memo No.1521 , C.B.C.L Paper No.112,1994.
[12] Bruce A. Draper Wendy S. Yambor and J. Ross Beveridge. Analyzing pca-based face recognition algorithms:Eigenvector selection and distance measures. In Second Workshop on Empirical Evaluation in Computer Vision, Dublin, Ireland, July 2000.
[13]D.L.Swets and J.Weng,“Using Discriminant Eigenfeatures for Image
Retrieval,”IEEE Trans.on PAMI,Vol.18,pp.831-836,1996
[14]H.Moon and P.J.Phillips,“Analysis of PCA-based Face Recognition Algorithms,”Empirical Evaluation Techniques in Computer Vision,K.W.Bowyer and P.J.Phillips,eds,IEEE Computer Society Press,Los Alamitos,CA,1998
[15]Wendy S.Yambor.Analysis of pca-based and fisher discriminant-
based image recognition algorithms.Master’s thesis, Colorado State University, 2000.
[16]On matching scores for LDA-based face verification J Kittler, Y P Li and J Matas Centre for Vision, Speech and Signal Processing University of Surrey, Guildford, Surrey GU2 7XH, UK J.Kittler@eim.surrey.ac.uk
[17]W.Zhao,R.Chellappa,and P.J.Phillips.Subspace linear discriminant analysis for face recognition.In UMD,1999
[18] The CSU Face Identification Evaluation System User’s Guide:
Version 5.0 Ross Beveridge, David Bolme, Marcio Teixeira and Bruce Draper Computer Science Department Colorado State University May 1, 2003
[19] Rowley, Baluja, Kanade，“Neural Network-Based Face Detection” PAMI, January 1998
[20]Penio S. Penev，Lawrence Sirovich ，“The Global Dimensionality of Face Space” , Proceedings of the Fourth IEEE International Conference on Automatic Face and Gesture Recognition 2000
[21]M.Gu and S.C.Eisenstat,“A Stable and Fast Algorithm for Updating the Singular Value Decomposition,”Research Report DCR/RR-966,Yale University,New Haven,CT,1994
[22]S.Chandrasekaran,B.S.Manjunath,Y.F.Wang,J.Winkeler,andH.Zhang,“An Eigen-space Update Algorithm for Image Analysis,”
Graphical Models and Image Processing,Vol.59,pp.321-332,1997
[23] Ross J. Micheals and Terry Boult. Efficient evaluation of classification and recognition systems. In IEEE Computer Vision and Pattern Recognition 2001, pages I:50–57, December 2001.
[24] W. Zhao, R. Chellappa, and A. Krishnaswamy. Discriminant analysis of principal components for face recognition. In In Wechsler, Philips, Bruce, Fogelman-Soulie, and Huang, editors, Face Recognition: From Theory to Applications, pages 73–85, 1998.
[25] J. Ross Beveridge, Kai She, Bruce Draper, and Geof H. Givens. A nonparametric statistical comparison of principal component and linear discriminant subspaces for face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
535 – 542, December 2001.
[26]徐士良，"计算机常用算法", 清华大学出版社，第二版，p137－p146
[27]王国印译，《Visual C++ 技术内幕》 清华大学出版社 ，1994
[28] 一种求解Fisher最佳鉴别矢量的新算法及人脸识别 郭跃飞 黄修武 杨静宇 中国图象图形学报Vol.4(A),No.2 1999年2月
[2] Kah-Kay Sung, Tomaso Poggio, "Example-based Learning for View-based Human Face Detection" , Massachusetts Institute of Technology, A.I. Memo No.1521 , C.B.C.L Paper No.112,1994
[3] 人脸检测研究进展 刘党辉 沈兰荪 文章编号1002－8331－（2003）28－0005－05 文献标识码 - A 中图分类号 TP391
[4] Viola P,Jones M.Fast and Robust Classification Using Asymmetric Adaboost and a Detector Cascade[R]. http://www.merl.com/people/viola/research/publications/2002.
[5] 边肇祺，张学工等编著，《模式识别》，清华大学出版社，2000年，第二版，p223－p227
[6] Liu K,ChengYQ,Yang JY.An efficient algorithm for Foley-Sammon optimal set of discriminant vectors byalgebraic method. International Journal of Pattern Recognition and Artificial intelligence,1982,6(5): 817～829.
[7] Viola P,Jones M.Robust Real Time Object Detection[C].In :Proc of IEEE Work shop on SCTV,2001.
[8] An Efficient LDA Algorithm for Face Recognition.Jie Yang, Hua Yu, William Kunz School of Computer Science Interactive Systems Laboratories Carnegie Mellon University Pittsburgh, PA 15213
[9] M.Turk andA. Pentland.Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):72–86, 1991.
[10]K.Etemad and T.Chellappa,“Discriminant Analysis for Recognition of Human Face Images,”Journal of the Optical Society of America A,Vol.14,pp.1724-1773,1997.
[11]Kah-Kay Sung, Tomaso Poggio, "Example-based Learning for View-based Human Face Detection" , Massachusetts Institute of Technology, A.I. Memo No.1521 , C.B.C.L Paper No.112,1994.
[12] Bruce A. Draper Wendy S. Yambor and J. Ross Beveridge. Analyzing pca-based face recognition algorithms:Eigenvector selection and distance measures. In Second Workshop on Empirical Evaluation in Computer Vision, Dublin, Ireland, July 2000.
[13]D.L.Swets and J.Weng,“Using Discriminant Eigenfeatures for Image
Retrieval,”IEEE Trans.on PAMI,Vol.18,pp.831-836,1996
[14]H.Moon and P.J.Phillips,“Analysis of PCA-based Face Recognition Algorithms,”Empirical Evaluation Techniques in Computer Vision,K.W.Bowyer and P.J.Phillips,eds,IEEE Computer Society Press,Los Alamitos,CA,1998
[15]Wendy S.Yambor.Analysis of pca-based and fisher discriminant-
based image recognition algorithms.Master’s thesis, Colorado State University, 2000.
[16]On matching scores for LDA-based face verification J Kittler, Y P Li and J Matas Centre for Vision, Speech and Signal Processing University of Surrey, Guildford, Surrey GU2 7XH, UK J.Kittler@eim.surrey.ac.uk
[17]W.Zhao,R.Chellappa,and P.J.Phillips.Subspace linear discriminant analysis for face recognition.In UMD,1999
[18] The CSU Face Identification Evaluation System User’s Guide:
Version 5.0 Ross Beveridge, David Bolme, Marcio Teixeira and Bruce Draper Computer Science Department Colorado State University May 1, 2003
[19] Rowley, Baluja, Kanade，“Neural Network-Based Face Detection” PAMI, January 1998
[20]Penio S. Penev，Lawrence Sirovich ，“The Global Dimensionality of Face Space” , Proceedings of the Fourth IEEE International Conference on Automatic Face and Gesture Recognition 2000
[21]M.Gu and S.C.Eisenstat,“A Stable and Fast Algorithm for Updating the Singular Value Decomposition,”Research Report DCR/RR-966,Yale University,New Haven,CT,1994
[22]S.Chandrasekaran,B.S.Manjunath,Y.F.Wang,J.Winkeler,andH.Zhang,“An Eigen-space Update Algorithm for Image Analysis,”
Graphical Models and Image Processing,Vol.59,pp.321-332,1997
[23] Ross J. Micheals and Terry Boult. Efficient evaluation of classification and recognition systems. In IEEE Computer Vision and Pattern Recognition 2001, pages I:50–57, December 2001.
[24] W. Zhao, R. Chellappa, and A. Krishnaswamy. Discriminant analysis of principal components for face recognition. In In Wechsler, Philips, Bruce, Fogelman-Soulie, and Huang, editors, Face Recognition: From Theory to Applications, pages 73–85, 1998.
[25] J. Ross Beveridge, Kai She, Bruce Draper, and Geof H. Givens. A nonparametric statistical comparison of principal component and linear discriminant subspaces for face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
535 – 542, December 2001.
[26]徐士良，"计算机常用算法", 清华大学出版社，第二版，p137－p146
[27]王国印译，《Visual C++ 技术内幕》 清华大学出版社 ，1994
[28] 一种求解Fisher最佳鉴别矢量的新算法及人脸识别 郭跃飞 黄修武 杨静宇 中国图象图形学报Vol.4(A),No.2 1999年2月