I 摘摘 要要 随着信息化社会的高速发展以及信息安全意识的普及，人脸检测与识别技术蓬勃发展并渗透进各行各业。人脸检测是识别的基础，依赖于所处的物理环境和检测装备的性能。本论文从物理环境中引起成像图片色彩偏差的光照条件着手，从机器视觉的角度修正偏差以实现最大可能的还原真实色彩，进而确定面部肤色区域，实现人脸检测。本论文的主要工作从以下四个方面进行阐述。 1、为了克服光照变化引起的机器视觉对图像颜色认知的影响，本文提出一种基于流形结构的色彩恒常算法。通过将一个宽的高斯模型与测试图像卷积提取背景光照，对参考光照主成分分析提取主要光照分量，结合流形结构将测试光转换为投射光照，最后从测试图像中滤去原始光，并覆盖投射光，实现背景光照的有效转换，并在肤色检测中得到验证。 2、在研究经典颜色恒常算法的基础上，从人眼视觉图心倾向性出发，加权平均视野内局部的亮度与色度信息，并结合无色差反射与最大响应点的优势，提出基于亮度加权的Gray- world 与 Retinex 融合的色彩恒常算法，能较好的实现色彩恒常处理。 3、 提出改进的双椭圆肤色模型， 把彩色图片从原彩色空间分别转换到 YCbCr 和 YCgCr空间后，考虑到色度对亮度的依赖性，利用非线性变化使色度 Cb、Cr、Cg 分量与亮度尽量分离，并拟合椭圆参数，逼近肤色在色度空间的聚类模型。 4、通过对二值肤色分割图像几何特征分析，实现人脸粗定位。Adaboost 算法通过引入Haar 矩形特征，构造弱分类器为强分类器，最终构造级联分类器,并结合色彩恒常算法，有效避免漏检，精确检测人脸。 关键字关键字：人脸检测；色彩恒常；双椭圆肤色模型；Adaboost PDF pdfFactory www.fineprint.cnII Abstract With the rapid development of information society , and popularity of information security awareness , face detection and recognition technology penetrate deep into all walks of life. Face detection is the basis of recognition，which depends on the physical environment and the performance of the testing equipment . This paper begin with light conditions which causes image color deviation. Update the bias to the true color from the perspective of machine vision , and then determine the facial skin color region . This paper expounds the following four aspects in turn: 1、In order to overcome the distraction of cogniting color caused by background illumination , this paper put forward a manifold based methodology for color constancy . Firstly, extracting illumination information with a wide Gaussian; Secondly, learning the manifold; Thirdly, projecting to the reference illumination; Finally, applying the projected illumination. This method works well in the convertion of the background illumination, and got verified in the skin color detection. 2、This paper put forward a new color constancy algorithm mixed with Gray- world and Retinex , equalize the local brightness and chrominance information in the human visual field ,which is uniting the superiority of non- aberration reflection and maximum response points. This method can better achieve color constancy . 3、Following the preprocessing of background color , this paper put forward a new modified double- ellipses skin color model. Separate illumination and chromaticity Cb、Cr、Cg using nonlinear variation after converting the origion picture from RGB to YCbCr and YCgCr color space in the view of the dependency . Approaching skin color clustering model by matching ellipse parameters. 4、 Achieve face coarse location by geometrical characteristic analysis with two- value splited image. Adaboost algorithm introduce Haar rectangle ferture , accumulate weak classifiers to strong classifiers, and finally construct cascade. It can detect the human face accurately and avoid leak detection effectively when combined with the proposed color constancy. Key words: Face Detection; Color Constancy; Double- ellipses Skin Color Model; Adaboost PDF pdfFactory www.fineprint.cnIII 目录目录 摘 要.I Abstract.II 目录.III 第一章 绪论.1 1.1 课题研究背景及其意义.1 1.2 研究现状.2 1.2.1 人脸检测在国内外的研究现状.2 1.2.2 颜色恒常在国内外的研究现状.3 1.3 论文主要研究内容和结构安排.6 1.3.1 论文研究内容.6 1.3.2 论文结构安排.8 第二章 基于色彩空间的肤色检测算法.9 2.1 颜色空间.9 2.1.1 RGB 颜色模型 .9 2.1.2 HSV 颜色模型.10 2.1.3 YCbCr 颜色模型.11 2