【作者】 王先基；

【导师】 庄镇泉； 郭立；

【作者基本信息】 中国科学技术大学 ， 电路与系统， 2007， 博士

【摘要】 自动人脸识别是一个典型的图像模式分析、理解与分类计算问题，它涉及到模式识别，图像处理，计算机视觉，统计学习，和认知科学等多个学科。自动人脸识别问题的深入研究和最终解决，可以极大的促进这些学科的成熟和发展。同时作为生物特征识别主要研究内容之一的人脸识别和认证技术在国家安全和公共安全领域的应用前景十分广阔。虽然经过近40年的发展，但是由于人脸容易受到光照、姿态和表情等因素的影响，要建立一个精确和鲁棒的自动人脸识别系统仍然是一个极具挑战性的课题。近年来基于统计学习的模式识别方法引起了极大的关注，在自动人脸识别领域也取得了很大的成功，使得自动人脸识别系统在速度和精度方面都得到了比较大的提高。代表算法包括Boosting算法，SVM和贝叶斯学习等。本论文就是针对统计学习算法在自动人脸识别中的各个环节中的应用展开的，论文的主要工作和创新成果如下：1．通过广泛的调研，对人脸识别的历史和现状进行了比较全面的综述本文首先对人脸识别研究的历史和发展现状作了回顾，之后对人脸识别中的一些主要算法作了比较详细地介绍，重点介绍了基于静态图片的人脸识别算法，另外对最近几年备受人们关注的基于视频的人脸识别研究也作了介绍。最后还介绍了目前人脸识别系统的评测情况及国内外主要的公用人脸数据库，在此基础上分析了当前人脸识别研究面临的挑战及可能的技术发展趋势。2．针对AdaBoost算法用于人脸检测特征选择时存在的问题，提出了一种基于代价敏感的AdaBoost算法的人脸检测方法目前人脸检测的主流方法是Viola和Jones在2001年提出了基于AdaBoost算法的人脸检测方法。该算法的一个缺点是训练分类器时，将两类分类错误(漏检一张人脸和误检一张人脸)平等对待，即漏检一张人脸和误检一张人脸的代价是一样的。因为在实际应用中人脸的存在和出现是个小概率事件，所以漏检一张人脸的代价应当比误检一张人脸的代价大。为此，本文提出了一种代价敏感的AdaBoost算法，此算法设定漏检一张人脸的代价比误检一张人脸的代价大，然后通过学习使得两类分类错误的代价最小。实验结果表明，此算法能取得更好的学习效果，提高了人脸检测率。3．研究了人脸识别中用AdaBoost算法挑选特征时存在的不对称性问题，提出了一种基于AsymBoost和Fisher线性判别分析的人脸识别方法本文详细分析了用AdaBoost算法在人脸识别中挑选特征时碰到的非对称问题，如正负样本数量相差悬殊和学习目标的不对称。提出用非对称AdaBoost算法(AsymBoost)来解决人脸识别中的不对称问题，以及采用Fisher线性判决分析对挑选出来的弱分类器的权重进行优化，使得综合后的分类器能最大化不同类别数据间的可分性。实验结果表明，采用AsymBoost算法和Fisher线性判决分析后分类器的识别效果有较大提高。4．研究了贝叶斯学习在人脸识别中的应用，提出了一种基于LBP特征和贝叶斯概率统计模型的姿态鲁棒的人脸识别算法本文分析了姿势变化对基于LBP特征的人脸识别算法的影响，提出用贝叶斯概率模型对姿势变化进行建模，使得基于LBP的人脸识别算法对姿势的变化更鲁棒。实验结果表明，相比原基于LBP特征的人脸识别算法，本文算法在一个更大的姿势变化范围内能取得更高的识别率。更多还原

【Abstract】 Automatic face recognition is a typical pattern analysis, understanding and classification problem, which is closely related to many disciplines such as pattern recognition, image processing, computer vision, statistical learning, and cognitive Psychology etc. The in-depth study and final settlement of AFR can greatly promote the maturity and development of these disciplines. While as one of the main research areas in Biometrics, face recognition and authentication techniques are believed having a great deal of potential applications in national security and public safety. Though face recognition has achieved great progress in the last 40 years, it is still a great challenge to build an automatic, high performance, high robust system for face recognition, due to the influence of illumination, pose and expression etc.In recent years, pattern recognition methods based on statistical learning have attached a great deal of attention which have achieved great success in AFR and greatly improved the speed and accuracy of AFR system. The representative algorithms include Boosting, SVM and Bayesian learning. This dissertation is just directed at the application of statistical learning algorithms in the various stages of AFR. The main work and innovation of this dissertation includes:1. Through extensive investigation and research, provided a thorough survey of the AFR history and the state-of-the-artAt first, this dissertation provided an overview of the history and the development status of AFR. Then, some main algorithms in face recognition are introduced, especially those based on static face images. In addition, the research on video-baed face recognition is also introduce, which has attached much attention in recent years. In the last, we survey the main public face databases and the performance evaluations protocols, based on which we also analyze the challeges and technical trends in AFR fields. 2. Against the existing problem of using AdaBoost for feature selection in face detection, proposed a face detection algorithms based on cost-sensitive AdaBoostCurrently, the main face detection method is the one based on AdaBoost algorithm proposed by Viola and Jones in 2001. The main disadvantage of the algorithm when training classifiers is that the two types of misclassification errors (having a face undetected and having a false alarm) are treated equally. Because in the real applications, the existing and emerging of faces is a small probability event, the cost caused by having a face undetectd is larger than that of having a false alarm. In this thesis, we propose a new AdaBoost called Cost-Sentitive AdaBoost, which sets the cost of false negative is larger than that of false positive and then seeks to minimize the total misclassification cost. The experimental results show that the proposed method can achieve better learning performance and improve the face detection rate.3. Against the existing asymmetry problems of using AdaBoost for feature selection in face recognition, proposed a new face recognition algorithm based on AsymBoost and Fisher linear discriminant analysisThis thesis analyzes in detail the asymmetries in face recognition when using AdaBoost for feature selection, such as the uneven distribution of the positive and negative samples and learning goal asymmetry. Then, we propose to use the asymmetric AdaBoost (AsymBoost) to address the asymmetries in face recognition, and adopt Fisher linear discriminant analysis (FLDA) to optimize the weights of the selected weak classifiers to maximize the separability between the data of different classes. The experimental results show that the performance of the final classifier trained by AsymBoost and FLDA is improved. 4. Studied the application of Bayesian learning in face recognition, proposed a robust pose-invariant face recognition based on LBP and Bayesian probabilistic modelThis thesis studies the influence of pose variation on the face recognition algorithms based on LBP, then proposed to Bayesian probabilistic model to model the pose variation to make the LBP face recognition algorithm be more robust to pose variation. The experimental results demonstrate that the proposed method achieved higher recognition rates in a wider range of pose changes compared to the original LBP face recognition method.更多还原