一种多视角人脸检测方法剖析.pdf

CN43- 1258/ T P 2011 33 1 ISSN 1007- 130X COMPUTER ENGINEERING SCIENCE Vol33, No1, 2011 : 1007- 130X(2011) 01-0132-0 * A Face Detection Method Based on Mult-i View , MA Ying-zhe, SUN Jin-guang (, 125105) ( School of Electronics an Information Engineering, Liaoning Technical University, Hulu ao 125105, China) : 目前的人脸检测方法多是针对正面人脸, 而对于多视角人脸检测还存在很大困难, 有效的方 法还不多本文 虑到人脸检测中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 法针对平面内旋转的问题, 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 灰度加权的主成分分析算法进行人脸的角度校正, 得到偏转校正后的人脸图像针对侧面人脸的问题, 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, 再对偏转校正后的人脸 图像进行人脸验证实验结果表明, 该方法可以对任意视角的人脸进行有效的检测, 且有较高的检测率 Abstract:The current method for the detection of human faces is mostly the frontal-view detection, w hile the mult-i view detection is still difficult, w ith few effective methods. Considering the rotation face and side face, this paper presents a mult-i view face detection method. For rotating on the plane, w e gen- erate a human skin model in the YCbCr color space. After treatment, the elliptic color area of face is de- termined. Then, using the gray weighting of the principal component analysis algorithm for angle cor- rection, w e obtain the correction face image. For the side face, through the face set of tw o directions of up or dow n and left or right, w e train the classifier, and combine them into a parallel connection classifi- er for validation. Experiments prove that this method is adaptive to different view point face and has a high face detection rate. : 人脸检测; 多视角; 肤色模型; 角度校正; 并联分类器 Key wor s:face detection; mult-i view ; skin model; angularity correction; parallel connection classifier oi:10. 39 9/ j. issn. 1007- 130X. 2011. 01. 02 : T P391