【作者】 宋加涛；

【导师】 刘济林； 池哲儒；

【作者基本信息】 浙江大学 ， 通信与信息系统， 2004， 博士

【摘要】 人脸自动识别是一种利用计算机分析人脸图像特征以实现人的身份验证的技术，是近二十年米图像处理、模式识别和计算机视觉领域中极富挑战性的研究课题之一。由于它在法庭举证、持卡人识别、视频监控等方面都具有巨大的应用价值，目前受到各国政府及其军事、安全、情报部门以及科研单位的广泛关注和高度重视。 一个完整的人脸识别系统由人脸检测、人脸图像预处理、人脸特征提取、图像间相似度计算等功能模块组成。多年来，人们对人脸识别技术中的许多问题都进行了深入的研究，并且已经研制出了不少有效的算法，然而，由于不同的人脸具有内在的相似性，而同一人脸的不同图像则常常因为姿势变化和表情变化而表现出巨大的差异性，并且人脸图像的质量还容易受到光照变化的影响，因此，现有的人脸识别技术仍然无法满足实际应用的需要。 本文重点研究人脸中双眼定位问题利强光照鲁棒性人脸识别算法开发问题。眼睛的位置是实现人脸图像几何归一化的必要条件，双眼的定位因而成为全自动人脸识别中非常关键的一环：而好的光照鲁棒性则能使人脸识别系统更富实用性。基于二值边缘图像，本文提出了一种新的人眼自动定位方法，以及一种具有较好光照鲁棒性的人脸识别方法。具体说来，本文主要做了以下三个方面的研究。 第一，提出了一种基于小波多分辨率分析的人脸二值边缘图像提取方法。该方法包括一次图像的高频重构过程、两次图像自适应二值化过程和一次后续的二值边缘图像去噪过程。定性和定量的评价表明，用该方法得到的二值边缘图像(BEI, Binary Edge Image)具有人脸部件完整、细节清晰、不同部件之间粘连少的优点，且具有很好的光照鲁棒性，因此，非常适合于人脸部件的分割和脸部关键特征点的提取。 其次，提出了一种基于BEI和亮度信息的人眼定位方法。该方法包括人脸区域提取、眼睛区域分割以及眼睛精确定位三个过程。为了提高眼睛检测和定位的成功率，该方法包含了一个人脸边界优化算法和一个多级人眼检测方案。另外，该方法将虹膜上的反射光点作为眼睛定位的重要线索，为此，该方法还包含了一个高鲁棒的反射光点自动检测算法。利用150个Bern图像和564个AR图像进行实验，分别获得了98.7％和96.6％的双眼定位成功率，表明该方法对眼睛定位过程中的视角变化、人脸表情变化和图像光照变化都具有很好的鲁棒性。进一步的分析和观察还表明，利用反射光点进行眼睛定位，不仅能获得很高的眼睛定位精度和成功率，而且能够实现对一些被头发部分遮住的眼睛的准确定位。 最后，提出一种基于二值模板匹配的人脸识别方法。该方法用由LAT(Locally Adaptive Threshold)算法获得的人脸二值边缘图像(BEM, Binary Edge Map)表征人脸，通过计算两幅BEM中重叠的前景像素数在其总的前景像素数中所占的比例来衡量两幅人脸图像的相似度。用AR和Yale人脸图像的实验表明，该方法能够以较快的速度获得比PCA法、灰度模板匹配法以及双修改Hausdorff距离方法更高的识别率，特别是对存有光照变化的图像。另外，本文还提出了一种进一步提高二值模板匹配识别性能的有效方法——在其决策过程中融入少量的灰度信息。实验表明，这种基于特征融合的扩展二值模极匹配法能同时获得较高的光照鲁棒性和表情鲁棒性，从而更加符合实际应用的需要。更多还原

【Abstract】 Automatic Face Recognition (AFR) is a technology for person authentication by using the digitized facial features. For the past two decades, it has become one of the most challenging research topics in the field of image processing, pattern recognition and computer vision. Because of its tremendous potential applications in law enforcement, security control, and video surveillance, AFR has attracted more and more attentions from many research institutes and government organizations including Departments in charge of defense, security, and information.A Fully Automatic Face Recognition System (FAFRS) consists of functions including face detection from an input image, face image pre-processing, facial feature extraction, and similarity measurement of two face images. These problems have been intensively investigated by researchers and many useful algorithms have been developed. Since faces of different subjects are often similar while face images from the same person often differ quite significantly due to pose and expression variations, and also because the quality of face images is affected greatly by lighting conditions, current face recognition systems cannot meet the requirements of many practical applications.This thesis focuses on the research of human eye localization and the development of face recognition algorithms with good robustness to illumination variations. The positions of two eyes are commonly used for the geometry normalization of a face image, eye locating is thus a very crucial step for the establishment of an FAFRS. The second research focus of this thesis was proposed because an FAFRS with good illumination robustness will be more applicable. Based on the binary edge image obtained using our proposed method, a novel eye localization method is presented in this thesis. Also based on the Binary Edge Map (BEM) obtained using the LAT (Locally Adaptive Threshold) algorithm, a new face recognition algorithm with better illumination robustness is presented. In particular, this thesis makes three main contributions detailed below.Firstly, a new face edge extraction method based on the multi-resolution property of Wavelet Transform (WT) is proposed. The method is composed of an image reconstruction step with high frequency components, two adaptive binarization steps and a noise removing step. The quantitative and qualitative evaluation show that face components, such as eyebrows, eyes, nose and mouth in the resulting Binary Edge Image (BEI) are all extracted with clear details and without touching neighboring face components in most cases. Furthermore, BEI is of good robustness to lighting changes. All these suggest that the BEI is suitable for face component segmentation and the extraction of some key feature points in the face image.Secondly, a novel method for the localization of human eyes is presented. The method consists of three steps, that is, face region extraction, eye region extraction, and finely locating of eyes. In order to improve the correct rate of eye locating, an algorithm for the refinement of face boundaries and a multi-level eye detection scheme are included in this method. In addition, thereflected light dots in the iris are used as an important cue for eye localization. Accordingly, an algorithm for the automatic detection of reflected light spots is given. Experimental results on a set of 150 Bern images and another set of 564 AR images show that correct eye locating rates of 98.7% and 96.6% have been achieved, respectively. The proposed eye detection method is also robust to the variations in views, lighting conditions and facial expressions. Furthermore, it is observed that by using reflected light dots for eye localization, those eyes partially covered by hair can be correctly located.Finally, a face recognition method based on binary template matching is proposed. This approach uses BEM’s for face representation. The similarity measure of face images is expressed as the ratio of the number of foreground pixels in the corresponding region to the total number of foreground pixe更多还原