人脸图像灰度分布统计分析与检测特征设计  

Statistic analysis of gray distribution on human face images and detection feature design

作　　者：欧凡[1] 刘冲[1] 

机构地区：[1]大连理工大学机械工程学院,辽宁大连116024

出　　处：《大连理工大学学报》2010年第4期528-534,共7页Journal of Dalian University of Technology

摘　　要：人脸图像的明暗视觉效果是面部五官稳定分布模式的外在表现,灰度分布模式可用做区分人脸与非人脸的依据.相对灰度比绝对灰度的抗光照干扰能力更强,为此以相对灰度均值和灰度标准差的比值衡量任一像素的明暗稳定程度,据此划分多个人脸明暗区域,将明暗区域配对组合为域灰度差检测特征,并基于优选出的28个域特征构建的42个隐节点的三层反馈神经网络检测分类器,对包含有图像质量恶劣的图像集进行测试,等误差率低于1.7%.测试结果表明所提方法有效可行,仅用28个特征就达到了与Viola算法数千个Haar类灰度差特征相仿的检测效果.Diverse visual aspects of gray texture on human face images are the extrinsic representations of intrinsic structure composed of facial organs. The common structure pattern of gray distribution can be used as a criterion reference pattern for discriminating processing in face detection. So statistic analysis of the regularity of gray distribution on human face images and detection feature design are dealt with. Instead of using original acquired gray intensities,the statistic analysis is based on the relative gray intensities,which are more robust. The occurrences of brightness and darkness of a pixel or a region are indexed with their specific ratio：the ratio of the mean intensity to the standard deviation of gray intensities. The whole face can be segmented into different gray scale regions based on the specific ratio distribution. A sub-region＇s intensity in brighter region can be coupled with another sub-region＇s intensity in darker region to form a gray difference feature. Based on the analysis,28 gray difference features are constructed and a 3-layer neural network face detection classifier with 42 hidden nodes is also built. A test face dataset including bad quality images is built and tested using the above classifier,and the equal error rate is under 1.7%. The experimental results show that the proposed approach is very effective using only 28 features and yielding comparable detection rate as Viola＇s approach with several thousand features.

关 键 词：人脸检测 分类熵 机器学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]