【作者】 张德锋；

【导师】 仲崇权；

【作者基本信息】 大连理工大学 ， 生物医学工程， 2009， 硕士

【摘要】 人脸检测是指将人脸从图像或视频中检测出来并提取面部特征的过程,它是人脸识别、特征定位的首要环节。同时在监测跟踪、出入检查、医疗领域也有重要应用价值。人脸检测受背景、光线及头部姿势等因素的影响而变的复杂。本文在对人脸检测技术进行深入学习和研究的基础上,主要对图像去噪声,检测特征选取和检测精度提高三方面问题展开研究。针对图像噪声问题,采用“参考白”算法、均值滤波算法对图像进行光线补偿和平滑处理,消除了图像中的光线和噪点干扰。采用中值法分离出视频背景,采用帧间差分算法提取出包含人脸的运动前景,去除了冗余视频背景,提高了检测速度。选取肤色作为人脸检测特征,提出了样本提取与Fisher准则结合检测肤色的方法,利用非线性变换改进了肤色在RGB与YC_bC_r色彩空间的聚类性,求得肤色在RGB与YC_bC_r色彩空间的样本高斯分布,计算出肤色似然图。利用Fisher准则分析肤色似然图,求出分割阈值,将图像二值化,划分二值化图连通区域得到肤色窗口。针对肤色检测精度不高问题,应用改进的Adaboost算法训练出由三个强分类器组成级联分类器,精检测肤色窗口。各强分类器由haar弱分类器加权构成。最后提出用图像横向积分与连通域结合法提取出眼睛特征,利用对像素R色彩通道分析提出嘴唇特征。结合改进的肤色提取与Adaboost算法,可以有效滤除干扰,并提高了检测速度,降低了误检率,保证了较高的人脸检测率。研究工作有一定理论意义和应用价值。更多还原

【Abstract】 Face Detection is a process that detecting the human faces from images or videos and distill the characters. It is a key point of Face recognition and Facial feature extract. At the same time, It has important applicable value in some places such like track, pass check and medical treatment.Face Detection is a complex problem because of it influenced by background, light and head’s posture. Based on the study on Face Detection technology, Noise Filter, Detection Character and Precision Improve are the main research directions in this paper.Compensate the picture’s light and then filter the interference color by using reference white arithmetic and average evaluation arithmetic for filter the noise. Median arithmetic is used to separate the background, Median difference arithmetic is used to separate moving information, filter the background and expedite the detection. Choose the complexion as detection character. bring forward a arithmetic that synthesize Stylebook Distill and Fisher judgment. Improve the complexion clustering in RGB and YC_bC_r color space and gained the complexion distributing. Account out the resemble image. Analyze the resemble image by fisher judgment and gained the binary image. Search the connect areas and get the complexion windows. Complexion Detection’s not high precision, training a series sort implement include three strong sort implements to detect the complexion windows subtly. Each strong sort implement is buildup with lots of weak sort implements. Finally detect the eyes character by the complexion windows’ landscape orientation integral. Detect the lips character by analyzing the R channels signal.By using complexion and Adaboost arithmetic can eliminate the noise signal effectively. It can both pledge the high detection rate and the low error judge rate and improve the system’s operation speed. The research has some theory and application value.更多还原