【作者】 李劼；

【导师】 周激流；

【作者基本信息】 四川大学 ， 通信与信息系统， 2004， 硕士

【摘要】 人脸识别是模式识别与计算机视觉、生物识别技术的交叉学科，而人脸检测是人脸识别系统的关键环节。根据生物识别领域内最新研究表明，非线性样本的处理和降维是人脸识别研究现今面临的两个主要问题。 核函数作为一种有效的处理非线性空间(可分／不可分)样本和迅速降维的理论和方法，随着支持向量机的普及，在近年来的模式识别领域得到了广泛的关注。将“核方法”与传统的特征提取和分类方法相结合，相继产生了许多新颖、有效的检测识别方法。本文主要研究内容是核函数的基础理论、算法性能改进以及在人脸检测中的应用。 本论文首先概述了人脸检测研究的发展现状。在此基础上，系统阐述了核函数理论的历史起源、基础理论、应用算法。针对基于核函数的FISHER判别(KFD)、基于核函数的主元分析法(KPCA)以及核函数的构造和参数选取作了一定的研究探讨。最后将训练好的核函数人脸检测机器应用于非线性空间人脸样本的检测，取得了较为满意的结果。 第一部分，通过阅读、学习大量中英文文献，概述了人脸检测技术研究的现状，对人脸识别系统中的传统特征提取、分类方法作了综述。针对人脸检测技术的非线性处理能力、降维能力、实时性、鲁棒性，比较了各种方法的优劣，提出了存在的疑难问题和本文的研究任务。 第二部分，从分析神经网络的局限性和疑难点入手，讨论了支持向量机取得成功的两大关键要素——凸二次规划和核函数，由此引出了统计学习理论和核函数理论。然后探讨了一系列核函数思想与传统的模式识别方法相结合的“核方法”，如核FISHER判别、核主元分析法、核感知器，研究了核函数的构造和参数的选取优化。通过MATLAB仿真试验实现了这些方法，并对各自的性能进行了比较;基于核函数理论和现有的人脸特征提取方法，针对样本的非线性这一人脸检测领域的重难点，提出了改进的核函数学习机器。 第三部分，基于核函数方法，提出了一个基于核函数理论、方法的人脸检测及验证算法。用改进的SvM实现了人脸检测，改进的KPCA实现了人脸验证。并且给出了仿真试验的数据、图片等结果，并对结果进行了分类、比较和推论。实验证明，基于核函数的人脸检测验证算法具有突出的小样本学习能力，对处理非线性空间对象和降维十分有效，识别精度和鲁棒性均令人满意。【关键词】核函数人脸检测主分量分析支持向量机生物识别更多还原

【Abstract】 Human face automatic recognition system is a cross subject combined with pattern recognition, computer vision and biometrics, in which human face detection is a key factor.It is generally believed that there is no more important problems than the nonlinearity of examples from actual space and dimensionality reduction of examples space during the course of the research of human face recognition. As a key factor of the success of the support vector machine, a more novel theory and more valid method to deal with such problems as nonlinearity of swatch and dimensionality reduction, kernel function theory has attracted sweeping researchers to apply it to existing pattern recognition fields and develop itIn the first part, the research actuality of human face detection technology is summarized. The traditional methods for feature extraction and classification are enumerated and compared with each other, which can not give attention to two or morethings such as nonlinearity and little examples learning.In the second part, beginning from the analysis about the limitation and artificiality of Artificial Neural Network, the support vector machine (SVM) is introduced and described, of which the convex quadratic programming and kernel function lead to the success. Then the statistic learning theory and the hypostasis of the kernel function are discussed. Then a series of kernel-methods based on thekernel function for classification and discriminat are studied and developed such as Kernel-based Principal Component Analysis (KPCA), Kernel-based Fisher Discriminat, of which the emphases are the principal, function constructing, parameters of the kernel functions. The arithmetic above are emulated by MATLAB. At last, a novel ameliorated human face detection algorithm based on kernel function is proposed.In the third part, a face detection and validation algorithm based on kernel function is proposed, in which the ameliorated SVM is used to perform face detection and a trained KPCA to face validation.In the last part, the results of the above-mentioned emulation experiment is provided, which are classified, compared and deduced It is proved that the human face recognition system based on kernel function have outstanding little swatch learning ability and very valid to deal with nonlinearity and dimensionality reduction which the common methods can not solve such as Linear Fisher Discriminant, model matching and Eigenface.更多还原