【作者】 马颖哲；

【导师】 孙劲光；

【作者基本信息】 辽宁工程技术大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 人脸检测是人脸识别等技术的研究基础,随着计算机技术的发展,人脸检测已经成为一个独立的研究课题,并且已经开始广泛应用到全新人机界面、基于内容的检索、基于目标的视频压缩、数字视频处理、视觉监测等许多领域。本文围绕多视角的人脸检测展开讨论。通过分析前人的人脸检测研究,发现多是对正面图像进行人脸检测,或是有限角度的人脸检测,而对于多视角的人脸检测,有效的方法并不多。本文考虑到人脸检测中平面内旋转人脸和侧面人脸两种多视角情况,提出了一种多视角人脸检测方法。本文针对平面内旋转问题,在YCbCr色彩空间内建立肤色模型,经过处理确定人脸椭圆区域,利用基于灰度加权的主元分析算法进行角度校正,得到偏转校正后的人脸图像。针对侧面人脸问题,提出一种二维空间划分方法对样本角度空间进行划分,通过上下和左右2个方向的人脸旋转样本库来训练分类器,然后组合成并联分类器对偏转校正后的人脸图像进行人脸验证。进而设计实现了一个多视角人脸检测系统,实验结果表明,本文构建的分类器在人脸检测的检测率和误检率上均有较好的效果,该检测方法可以对多视角的人脸进行有效的检测。更多还原

【Abstract】 Face detection is the basis of technologies like face recognition, with the development of Computer technology, face detection has become a independent research subject. Face detection has been used in many fields such as human computer inter face, searches based on content, visions inspection and so on.This essay is centered of the constitution format of Multi-view in face detection. By analyzing the classifier built by former people, we find out that the current detection of human faces is mostly the frontal-view detection, and the view is limited, while the multi-view detection is still difficult, with few effective methods. Considering the rotate face and side face, this paper presents a multi-view face detection method. For rotating on the plane, we generate a human skin model in YCbCr color space. After treatment, elliptic color area of face is determined. Then, using the gray weighting of the principal component analysis algorithm for angle correction, and obtain the correction face image. For side face, through the face set of two directions of up or down and left or right, we train classifier, and combine them into parallel connection classifier for validation.This essay designed and implemented a Multi-view face detection system, which through comparison showed that the improved classifier have a better result both on detection rate and false positive rate. Experiments prove that this method is adaptive to different views face having a high face detection rate.更多还原