【作者】 陈思；

【导师】 杜太行；

【作者基本信息】 河北工业大学 ， 检测技术与自动化装置， 2007， 硕士

【摘要】 人脸研究,主要包括人脸检测、人脸跟踪、人脸识别以及衍生出来的姿态和表情分析等几个主要领域,其中人脸检测是人脸信息处理领域中的一项关键技术。最初的人脸研究主要集中在人脸识别领域,而且早期的人脸识别算法都是在认为已经得到了一个正面人脸或者人脸很容易获得的前提下进行的。但是随着人脸应用范围的不断扩大和开发实际系统需求的不断提高,这种假设下的研究不再能满足需求。人脸检测开始作为独立的研究内容发展起来。本文主要研究人脸的动态检测与姿态检测。达到能够在动态环境下,准确检测出人脸的位置,并从一系列各方向的人脸姿态中,快速检测出最正向的人脸姿态,用于后续的人脸的识别工作,提高识别的准确性。本文首先归纳和分析了当前人脸检测的典型算法,综合视频图像运动信息和彩色信息,通过运动检测限制搜索范围,然后提出了一个由粗到细的多阶段的筛选与验证方法,以定位背景中数目、大小、位置均未知的人脸。该方法首先分析、比较了肤色在各色彩空间中的聚类性,然后再在YCbCr色彩空间中建立肤色模型,对肤色进行了相似度分割和二值化处理。由于噪声无所不在,我们对二值化后的图像进行了基于数学形态学的滤波处理,之后将此结果作为候选人脸区域输出。在筛选与验证人脸阶段,本文首先根据区域面积占有率对人脸进行了初步筛选,再利用欧拉数和椭圆面积准则的方法,对筛选后的候选区域进行了再次筛选与验证,得出了人脸检测结果。姿态的判定选择对所检测出的人脸区域采用基于YCbCr空间的方法来找到似眼和似嘴区域,再通过先验知识去除假区域,找到眼及嘴的三个中心点,确定基于三点的三角形,我们称之为姿态三角形,当其满足一定要求时视其为正向姿态,保存正向姿态人脸图片,实现了课题的最终要求。本文的最后利用Visual C ++编程工具对该算法进行了实现,在实验室环境相对稳定,系统参数设定适宜,摄像机安放位置合适情况下,实验证明该方法可以有效地应用于多人脸、不同尺度和复杂背景的情况,具有良好的检测效果。但是由于人脸自身的复杂度、以及人脸中的许多遮挡,该系统还是存在一些漏检和误检,但总的来说,该系统仍具有一定的应用性。更多还原

【Abstract】 Research efforts in face processing include face detection, face tracking, face recognition as well as derivative analysis of pose and expression. Face detection is a key technology of the filed of face information processing. In the beginning, research efforts in face processing chiefly fastened its attention on face recognition, and early face recognition algorithms are based on having a frontal face or faces’getting easily. But with faces’application spreading and practical systems’demand improving, this kind of research on this condition can’t satisfy requirements any more. Face detection develops as an alone research.This paper mainly researches in the dynamic face detection and the face pose detection. To achieve the purpose that in dynamic context detect the position of face, and from all poses of faces find the most positive face to use in the face identification. That can improve the veracity in face identification.In this dissertation, we first sum up and analyze current typical algorithms on face detection, and then present a multistage detection and conformed method from coarse to fine in order to find faces in the complicated background. This method is discussed on details as follows. First, the algorithm analyzes and compares complexion’s clustering in the different color space, and then establishes a skin model based on the color space of YCbCr. Using this model, the complexion is segmented and two-value processed. On account of noise’s immanence, we can get backup face area using the filter based on mathematical morphology. In the selective and conformed stage,we utilize faces’geometrical characters to roughly choose between these backup areas. Finally we use Euler number and fair model to repeatedly choose and verify these regions,and to get the end results.The determinacy of face pose is based on the eyes and the mouth areas which are detected by the method based on the color space of YCbCr . Then ,delete the false area,find the middle of eyes and mouth area. The three points joint into a triangle. When the triangle meets some rules which makes face pose positive, save that picture as the result.In the last part of the dissertation ,we implement the algorithm using Visual C++. In the case of that the lab is steady ,the parameter of the system is proper, the camera is still and in the perfect place, experimental results show that the method is effective in many conditions including many faces’different scales and complex background. But due to face’s complexity by itself ,there are some missing faces and false faces. The system is practical to a certain extent.更多还原