人脸的几何特征提取与查询

【作者】 黄雪琳；

【导师】 蔡灿辉；

【作者基本信息】 华侨大学 ， 计算机应用技术， 2004， 硕士

【摘要】 人脸识别在诸如身份识别和公安档案等方面有着十分重要的应用价值。进入90年代以来，国内外研究人脸识别的方法层出不穷，其中包括基于几何特征的方法、基于特征脸的方法、基于弹性模型的方法、基于局部特征的方法、和神经网络方法等等。本文论述了在微机上人脸图像自动识别系统的实现，采用的是基于几何特征的方法。具体包括人脸图像的预处理、人脸图像中眼睛、鼻子、嘴巴各部位的自动定位、人脸的特征提取以及人脸图像的识别等几大功能。人脸图像的预处理包括图像的灰度变换、二值化处理及图像的归一化。在二值化处理时，我们采用的是既可以优化图像质量，又可以完成对图像的二值化处理的二次边缘提取算法。图像的归一化包括图像裁剪和缩放，从而实现了图像的位置校准。在人脸的特征提取方面，本文选取了人眼、鼻子、嘴的大小和位置等十五个特征值，从而使识别的误差大大减小。该算法简单易行而且对人脸的尺寸、脸形、光照、背景复杂性等没有限制，适用于不同质量的图像。 在进行人脸识别时，先求待识别人脸与数据库中人脸图像的Euclidean距离，得出误差最小的前八幅图像，再分别求出待识别人脸与数据库中这八幅人脸图像的左眼睛(含眉毛)、右眼睛(含眉毛)、嘴巴所对应三个点集的最小距离。然后取这三个最小距离中的最大者作为Hausdorff距离进行匹配，所得误差最小者即为目标记录。更多还原

【Abstract】 The face recognition may be applied in many fields such as identity recognition and public security document management. Since nineties, the methods of studying face recognition have emerged in endlessly, including geometry feature based method, characteristic face method, stretch model based method, local feature method and NN. This paper is about the realization of the automatic recognition system on facial images on computer. The system is based on geometry feature method. The system includes pre-processing of facial images, automatic location of the eye, nose and mouth ,and the character abstraction and recognition of face images. The pre-processing includes grey transformation binalization and unification. In binarization, we use the algorithm of the second brim abstraction with which we not only optimize the quality of the images but also finish the binarization. The unification, which realizes the position calibration of the images, includes cropping and zooming. In character abstraction of facial images, we choose fifteen characters which includes the size and position of eyes, nose, mouth, so that the error of the recognition is greatly reduced. The algorithm is simple and have no restrict to the size, shape, light and background of the face.When the face recognition is proceeding,we first calculate the Euclidean distance of the facial image to be recognited and the facial images in database,and get the front eight images of the least error margin resemble,and then calculate the minimum distance of the three points gathering respectively of the left eye(contain the eyebrow),right eye(contain the eyebrow) and mouth between the facial image to be recognited and those eight images.Taking the biggest value of these three minimum distances as Hausdorff distance to match.the target record is the least error margin record.更多还原