ABSTRACT
Face Recognition is the hot research topic in today’s field of Pattern Recognition, which is the important method of the identification identity. Because Face Detection is the premise and basis of Face Recognition, we have to correctly position out the place of faces in images or videos firstly, and then the effective recognition can be accomplished. Face Detection in the general target detection field has its important research significance.
In the first three chapters of this paper, we mainly introduced the Face Detection based on learning process, and this raised the Face Detection algorithm based on study of the representative-Face Detection method based on AdaBoost. AdaBoost algorithm was put forward in 1995, a rapid Face Detection algorithm, which was a milestone in the process of Face Detection field. This algorithm according to the feedback of the Weak-Learner, adaptively adjusts the assumption of the error rate, and from this, detection accuracy can get improved a lot without the decrease of efficiency.
In the fourth chapter, we introduced two important factors influencing AdaBoost training algorithm: haar rectangle feature and Integral Images. At the same time, we expanded haar rectangle feature from 5 to 15, which made it tend to be perfect.
In chapter 5, we introduced the training and cascade of classifier and the detailed process of AdaBoost algorithm. We improved the traditional Weak-Learner, and made a weak classifier composed of a big threshold and a small threshold such that the minimum numbers of samples were misclassified.
In chapter 6, we introduced the realization of AdaBoost algorithm based on OpenCV, and it focused on the training method of face classifiers.
At the end of the paper, we summarized the design plan, and made prospect of the future research direction of face detection.
Keywords：Face Detection; Integral Image; haar feature; AdaBoost; Cascade Classifiers
目 录
1 人脸检测概述 .................................................................................................................................................. 1
1.1 人脸检测概念 ....................................................................................................................................... 1 1.2 人脸检测的发展历史 ........................................................................................................................... 1 1.3 人脸检测的难点与研究方法 ............................................................................................................... 3 2 基于学习的人脸检测流程 .............................................................................................................................. 6
2.1 分类器的离线学习 ............................................................................................................................... 7
2.1.1 图像对齐 .................................................................................................................................... 7 2.1.2 预处理 ........................................................................................................................................ 8 2.1.3 特征提取 .................................................................................................................................. 10 2.1.4 分类器的学习 .......................................................................................................................... 10 2.2 目标的在线检测 ................................................................................................................................. 10 2.3 人脸检测方法的性能评价 ................................................................................................................. 12
2.3.1 人脸图像数据库 ...................................................................................................................... 12 2.3.2 性能评价方法 .......................................................................................................................... 13
3 基于AdaBoost的人脸检测方法概述 .......................................................................................................... 16 4 矩形特征与积分图 ........................................................................................................................................ 17
4.1 引言 ..................................................................................................................................................... 17 4.2 haar矩形特征 ...................................................................................................................................... 18
4.2.1 概述 .......................................................................................................................................... 18 4.2.2 特征模板与模板特征值 .......................................................................................................... 18 4.2.3 haar特征的扩展 ....................................................................................................................... 19 4.3 积分图Integral Images ....................................................................................................................... 21
4.3.1 积分图的概念 .......................................................................................................................... 21 4.3.2 基于积分图的像素值计算 ...................................................................................................... 22 4.3.3 基于积分图的特征值计算 ...................................................................................................... 23
5 分类器训练及AdaBoost算法 ...................................................................................................................... 24
5.1 弱分类器Weak Classifier ................................................................................................................... 25 5.2 AdaBoost算法训练强分类器 ............................................................................................................. 26
5.2.1 算法描述 .................................................................................................................................. 26 5.2.2 算法流程图 .............................................................................................................................. 27 5.3 级联强分类器结构 ............................................................................................................................. 29 6 基于OpenCV的程序实现 ............................................................................................................................ 30
6.1 OpenCV简介 ....................................................................................................................................... 30 6.2 训练样本集与测试集 ......................................................................................................................... 30 6.3 基于OpenCV的分类器实现 ............................................................................................................. 31
6.3.1 样本描述文件创建 .................................................................................................................. 31 6.3.2 训练分类器 .............................................................................................................................. 32 6.4 检测结果 ............................................................................................................................................. 33
6.4.1 检测结果对比 .......................................................................................................................... 33 6.4.2 基于AdaBoost算法的更多检测结果 .................................................................................... 34 6.4.3 基于AdaBoost算法的视频检测结果 .................................................................................... 37
7 结论与对未来的展望 .................................................................................................................................... 38 参考文献 ............................................................................................................................................................ 40 英文原文 ............................................................................................................................................................ 41 中文译文 ............................................................................................................................................................ 50 致 谢 ................................................................................................................................................................ 57
图像目录
图1.1 人脸的遮挡、不同表情和图像的旋转会影响到人脸的检测 ...................................................... 5 图1.2 不同角度人脸带来的检测结果也有可能不同 .............................................................................. 5
图2.1 基于学习的人脸检测流程 .............................................................................................................. 7 图2.2 目标检测的金字塔 ......................................................................................................................... 11 图2.3 检测结果的融合 ............................................................................................................................. 11 图2.4 典型的正面人脸图像数据库中的人脸图像 ................................................................................ 13 图2.5 左侧为测试图像，右侧为检测结果。 ........................................................................................ 14 图2.6 MIT+CMU测试集的人脸检测ROC曲线 ................................................................................... 15
图3.1 AdaBoost人脸检测器结构 ............................................................................................................ 16 图3.2 人脸检测系统流程 ........................................................................................................................ 17 图3.3 分类器级联结构 ............................................................................................................................ 17
图4.1 矩形特征在人脸上的特征匹配 .................................................................................................... 18 图4.2 竖直和45倾斜的矩形的例子 ..................................................................................................... 19 图4.3 扩展的haar特征集合 ................................................................................................................... 20 图4.5 灰度图像积分图算法示意图 ........................................................................................................ 22 图4.4 ―积分图‖与积分的类比 ................................................................................................................. 21 图4.6 区域D的像素和可以用积分图计算 ........................................................................................... 23 图4.7 矩形特征的特征值计算 ................................................................................................................ 24
图5.1 人脸检测中强弱分类器与AdaBoost算法 .................................................................................. 25 图5.2 弱分类器中的阈值选取 ................................................................................................................ 25 图5.3 算法流程图 .................................................................................................................................... 28
图6.1 MIT部分人脸样本 ......................................................................................................................... 31 图6.2 MIT部分非人脸样本 ..................................................................................................................... 31 图6.3 测试结果对比1 ............................................................................................................................. 33 图6.4 测试结果对比2 ............................................................................................................................. 34 图6.5 测试结果对比3 ............................................................................................................................. 34 图6.6 一张人脸 检出1 漏检0 虚警2 ................................................................................................. 34 图6.7 一张有遮挡的人脸 检出1 漏检0 虚警0 ................................................................................. 35 图6.8 四张有遮挡的人脸 检出4 漏检0 虚警2 ................................................................................. 35 图6.9 五张人脸 检出5 漏检0 虚警0 ................................................................................................. 35 图6.10 五张远近不同的人脸 检出5 漏检0 虚警3 ........................................................................... 36 图6.11 三张人群中模糊的人脸 检出3 漏检0 虚警1 ....................................................................... 36 图6.12 八张人脸 检出8 漏检0 虚警1 ............................................................................................... 37 图6.13 一张人脸 检出1 漏检0 虚警0 ............................................................................................... 37 图6.14 四张人脸 检出3 漏检1 虚警2 ............................................................................................... 38 图6.15 光线昏暗且有遮挡的两张人脸 检出2 漏检0 虚警1 ........................................................... 38
?
1 人脸检测概述
1.1 人脸检测概念
长期以来，计算机就好像一个盲人，需要被动地接受由键盘、文件输入的信息，而不能主动从这个世界获取信息并自主处理。人们为了让计算机看到这个世界并主动从这个世界寻找信息，发展了机器视觉；为了让计算机自主处理和判断所得到的信息，发展了人工智能科学。人们梦想，终有一天，人机之间的交流可以像人与人之间的交流一样畅通和友好。
迄今为止，机器视觉的发展已经历了一个漫长的过程。经过研究者们的不懈努力，新的信息技术和媒体手段的出现，使得更加有效和友好的人机交互方式得到了发展，新型的人机交互将不再依赖传统的输入设备。而且，计算性价比的提高和视频获取成本的降低，使得计算机视觉系统能够向桌面级和嵌入式系统发展，这意味着计算机视觉系统能够安装在一切电子系统之中。相信在不久的将来，拥有高级视觉系统的智能电子产品会给我们的生活带来更大的方便。
人脸检测(face detection, FD)，也称自动人脸检测(automated face detection, AFD)，是近年来得到广泛关注的研究热点，其相关技术的应用范围也在不断扩展，人脸检测技术的出现与飞跃，为机器视觉在人机交互领域的长足进步奠定了坚实的基础。
人脸检测是指在输入图像中判断是否存在人脸区域，并进一步确定人脸的位置、大小和姿态等信息。人脸检测是人脸识别、表情检测和性别检测等人脸信息处理问题中涉及的一项关键技术，同时也是通用目标检测问题最具代表性的特例，近年来在模式检测与计算机视觉领域受到普遍重视。
1.2 人脸检测的发展历史
人脸检测是一个被广泛研究着的热门领域，大量的研究论文层出不穷。为了更好地对人脸检测研究的历史和现状进行介绍，我们将按照研究内容、技术方法等方面的特点大体划分为三个时间阶段来做介绍。
第一阶段（1964-1990年）：
这一阶段人脸检测通常只是作为一个一般性的模式识别问题被研究，所采用的主要技术方案是基于人脸几何结构特征（Geometric feature based）的方法，这集中体现在人们对于剪影（Profile）的研究上，人们对面部剪影曲线的结构特征提取与分析方面进行了大量研究。人工神经网络也一度曾经被研究人员用于人脸检测问题中。较早从事AFR研究的研究人员除了Bledsoe外还有Goldstein，Harmon以及Kanade等。Kanade博士于1973年在京都大学完成了第一篇AFR方面的博士论