【作者】 程勇光；

【导师】 郝晓莉；

【作者基本信息】 北京交通大学 ， 电路与系统， 2008， 硕士

【摘要】 人脸检测需要综合应用图像处理、模式识别、计算机科学等知识才能达到预期的效果,是一项极有挑战性的研究。本文从实用性和易于实现的角度出发,研究了基于肤色分割和面部几何特征提取的人脸检测方法。首先,研究了利用肤色信息从背景中分割出脸部候选区域的算法。在YCbCr彩色空间内,对光照偏离较大的彩色图像进行光线补偿;然后,分析了两种肤色分割方法:一种是根据肤色的Cb和Cr呈高斯分布的特性,利用高斯模型对肤色区域进行分割,另外一种是在非线性变换后的YCbCr彩色空间中基于椭圆肤色模型的方法进行肤色分割;对分割出来的区域进行形态学处理,从而确定人脸侯选区域。前一种肤色模型的色度是独立于亮度的,后者则考虑了亮度变化对色度的非线性影响,因此,后者能更精确地检测出肤色区域。肤色分割得到的候选区域中,有的是人脸区域,有的是非人脸区域。因此,提出利用面部几何特征从候选区域中去除非人脸区域,确定人脸位置。利用人脸面部构造产生的灰度特性提取眼睛,利用嘴唇的色度特征分割出嘴巴,进而根据眼睛和嘴巴构成三角形模板的特性,精确定位人脸在图像中的位置。本文采用Matlab研究该算法。实验结果表明,这种结合肤色和面部几何特征的算法,能够对人脸进行较快速和准确的定位,并且结果比较稳定可靠。更多还原

【Abstract】 Human face detection is necessary for face recognition as its first step to locate faces in images. Face detection has been intensively researched and has developed as an independent field.A face detection approach based on the skin color and facial geometric features is presented for its practicability and facility in this paper.We detect skin regions based on color information is researched in the former half of the paper. The first step is preprocessing images, such as filtering the images with low-pass filter, lighting compensation with the different stretching methods in different gray levels. We studied two methods to find face candidates. One is building a Gaussian model of skin-color to segment the skin area, and the other is the method of nonlinear color transformation in YCbCr. The latter method is more effective for varying lighting conditions.Some of segmented regions are faces, some of them are not. We use facial geometric features to decide which ones are real faces. Eye model based on gray-level information and mouth model based on chrominance are constructed respectively, then a triangle model based two eyes and a mouth is employed to verifying each candidate face.The experiment results show that this algorithm is effective to detect faces for color images with clear faces.更多还原