【作者】 邵平；

【导师】 杨路明；

【作者基本信息】 中南大学 ， 计算机应用技术， 2007， 博士

【摘要】 人脸检测在人机交互、基于内容的检索、数字视频压缩、视频监控等许多领域具有广阔的应用前景,是各种人脸处理系统最为基础而又十分重要的技术环节,是近年来国内外研究的一个热点。然而目前复杂背景下的多姿态人脸检测还存在很大困难,有效的方法还不多。本文主要研究了复杂背景下的多姿态人脸检测及相关的快速算法。首先以提高人脸检测速度为主线,研究了复杂背景下人脸检测的各个环节,提出了一些快速算法;然后分别设计了灰度图像和彩色图像的快速多姿态人脸检测方案。具体说来,做了以下工作:第一,针对人脸检测的各个环节,提出了一些相应的快速算法。针对模板匹配前的灰度分布标准化较为耗时的问题,提出了一种基于广义积分图像的灰度分布标准化快速算法,使图像窗口灰度均值和方差的计算时间大大减少;为综合利用图像梯度特征及梯度方向特征进行人脸检测,采用多方向Kirsch边缘检测算子,并就此提出了一种基于模板分解和积分图像的快速Kirsch边缘检测算法,以解决多方向边缘检测耗时问题;积分投影和方差投影是进行人脸检测和器官定位的常用方法,就此提出了一种基于广义行-列积分图像的快速投影算法,有效地降低了图像窗口行(或列)积分投影和方差投影的计算量;另外,针对多模板匹配的计算效率问题,提出了基于广义积分图像的快速多模板匹配算法;针对模板匹配时存在掩膜的情形,提出了基于广义掩膜积分图像的快速模板匹配算法,并说明了其在多姿态模板匹配中的应用。第二,针对复杂背景下灰度图像中的多姿态人脸检测问题,提出了一种基于知识模型和模板的快速多姿态人脸检测算法。即首先从原始图像中提取人脸器官梯度图,并建立多姿态知识模型和多姿态模板;然后以多姿态知识模型和知识规则进行人脸粗检,以多姿态模板匹配进行人脸细检,从而得到人脸在图像中的位置和大小信息,并通过眼嘴重心构成的三角形估计人脸的粗略姿态。多姿态知识模型中任意矩形区域的器官梯度特征点快速求和利用积分图像来实现;多姿态模板与图像窗口的局部和整体匹配应用本文提出的快速多姿态模板匹配算法来实现。第三,针对复杂背景下彩色图像中的多姿态人脸检测问题,提出了一种基于多阈值特征融合的快速多姿态人脸检测算法。即首先从原始图像中,根据人脸器官梯度特征和Kirsch边缘检测算子提取多阈值器官梯度图和梯度方向图,并根据人脸肤色特征提取双阈值肤色图,根据亮度信息提取灰度特征图;然后建立特征融合模型,并应用多姿态知识模型和多姿态模板实现人脸检测。人脸检测过程中,采用了由粗至精的检测策略,并在各个环节应用相关的快速算法,以提高人脸检测的速度;融合了能在复杂背景下区分人脸的多种特征,以提高人脸检测的准确性,还融合了同一特征在不同阈值下的信息,以减小人脸漏检的可能性。另外,根据人眼能不依赖肤色从灰度图像中轻松地分辨出人脸的事实,允许梯度特征具有明显人脸模式的窗口跳过肤色检测直接进入下一级人脸分类器,以减少肤色失真造成的漏检。更多还原

【Abstract】 Face detection can be widely used in many fields such as human-machine interaction, content-based retrieval, digital video compression and video surveillance. Face detection, which is also a hot topic in domestic and foreign research areas in recent years, is the most basic and very important technique for all sorts of face processing systems. However, there are still difficulties for multi-pose face detection under complex background, and only a few effective methods can be found at present.This dissertation focuses on multi-pose face detection under complex background and some correlative fast algorithms. Following the main clue of improving face detection speed, this dissertation has studied each tache of multi-pose face detection under complex background, and proposed some fast algorithms. Then the schemes for fast multi-pose face detection in gray and color image are designed respectively. Actually, this dissertation makes some detail contributions as follow:Firstly, some correlative fast algorithms are presented, aiming at each tache of face detection. As it is very time consuming for grayscale distribution normalization before template matching, a fast algorithm is presented for grayscale distribution normalization based on extended integral image. This algorithm reduces greatly the time for calculating the grayscale average and variance of image windows. In order to make full use of the image grads and grads direction feature for face detection, multi-direction edge detection operators of Kirsch is used. Therefore, a fast algorithm of Kirsch edge detection based on templates decomposition and integral image is proposed for solving the time consuming problem of multi-direction edge detection. Since face detection and organ locating with integral projection and variance projection are in common use, a fast projection algorithm based on extended row-column integral image is presented, and then the calculating time is reduced effectively for the row (or column) integral projection and variance projection of image windows. In addition, aimed at improving the efficiency of multi-template matching, a fast multi-template matching algorithm based on extended integral image is presented. For application of template matching with mask, a fast template matching algorithm based on extended integral image with mask is proposed, and the detailed application in multi-pose face template matching is illustrated.Secondly, aimed at solving the problem of detecting multi-pose face in gray image under complex background, a fast face detection algorithm is proposed based on multi-pose knowledge models and templates. At first, a face organ grads image is extracted from the original image, and multi-pose knowledge models and the multi-pose face templates are set up. Then coarse face detection is implemented by multi-pose knowledge models and correlative rules, and fine face detection is done by multi-pose face template matching. Finally the position and size of each face are captured in image, and the coarse pose of each face is estimated by the triangle with center of gravity of eyes and mouth. The organ grads feature pixels in random rectangle areas is summed up rapidly by integral image in multi-pose knowledge models. The local and whole matching between multi-pose templates and image windows is implemented by fast multi-pose templates matching algorithm proposed in this dissertation.Thirdly, aimed at solving the problem of detecting multi-pose face in color image under complex background, a fast multi-pose face detection algorithm based on multi-threshold feature fusion is presented. Multi-threshold organ grads image and grads direction image are extracted by face organ grads feature and Kirsch edge detection operators from the original image, double-threshold skin image is extracted by face skin feature, and gray feature image is extracted by lightness information. Then the feature fusion models are set up, and multi-pose knowledge models and multi-pose face templates are used to accomplish face detection. In face detection processing, coarse-to-fine strategy and correlative fast algorithms in each tache are used to improve the detection speed, several features for distinguishing face under complex background are fused to improve the detection accuracy, and the missed faces are reduced by fusing information of the same feature in different thresholds. In addition, based on the fact that human eyes can detect face easily from gray image without skin color information, windows with obvious grads feature of face pattern are allowed to enter the next face classifier without skin color detection, so the missed faces with skin color distortion can be reduced.更多还原