【作者】 胡文静；

【导师】 刘锦高；

【作者基本信息】 华东师范大学 ， 无线电物理， 2006， 博士

【摘要】 自动人脸识别技术(AFR)是一项极具挑战性的前沿研究课题。它试图通过计算机分析人脸图像并从中提取有效识别信息，达到辨认人员身份的目的。对AFR技术的研究不仅具有重大的理论和学术研究意义，而且具有潜在的巨大应用价值。 经过近几十年特别是近几年来的研究，自动人脸识别技术已经取得了长足发展，用于人脸识别商业系统已经面市，但对应用条件的限制相当严格；在非理想可控情况下的自动人脸识别技术还远未达到实用化的程度，有很多研究工作要做。 本文作为上海市应用材料科技国际合作共同计划(上海市科委AM基金)项目(《基于ARM和RFID芯片的自组织安全监控系统的研制》编号：0512)的主要研究内容之一，从构建自动人脸识别系统需要解决的若干关键问题入手，重点探讨了实时人脸检测与跟踪、面部关键特征定位、高效的人脸特征描述、鲁棒的人脸识别分类器及自动人脸识别系统设计等问题。 1、提出了结合肤色校验的Haar特征级联分类器实时人脸检测算法(SCC-HCC)和基于人脸约束的人脸实时跟踪算法(AM-CamShift) 人脸检测是自动人脸识别系统首先需要解决的关键问题。Viola于2001年提出的基于Haar特征级联强分类器的人脸检测算法，通过抽取人脸的Haar特征训练分类器，达到人脸检测的目的，但由于其仅仅利用了人脸的灰度信息，没有考虑人脸的肤色分布，因而对复杂背景中类人脸结构的物体对象区分的鲁棒性较差。鉴于此，论文第三章提出了基于肤色模型校验和Haar特征级联强分类器的快速人脸检测算法(SCC-HCC)。 人脸跟踪是基于视频的人脸识别、视频监控等典型应用中必不可少的环节，CamShift算法对于目标物体的跟踪具有较强的鲁棒性，但其存在跟踪窗口(Tracking Window)必须通过手工标定的缺陷，而且对背景中类肤色区域的鲁棒性欠佳。我们在CamShift算法的基础上提出了基于人脸约束的实时跟踪算法(AM-CamShift)，实现了跟踪窗口自动标定及多目标的快速自动跟踪，有效提高了对背景中类肤色区域的鲁棒性。 2、针对传统线性判别分析法存在的小样本问题(SSS)，通过调整Fisher判别准则，实现了自适应线性判别分析(A-LDA)算法，提出了基于A-LDA算法的分类判决准则及相应的人脸识别方法更多还原

【Abstract】 Automatic Face Recognition (AFR) is challenging in image processing and analyzing. By AFR, we mean that people attempt to endow computers with the ability to analyze the human face image, to extract the valid individual information, and to identify him/her. Such a type of theory and technology is not only greatly desired in the theoretical research but also has significant potential in applications.After the nearly tens years’ research, AFR technology has already obtained considerable advances especially in the past several years. The existing AFR commercial systems can basically meet the application need under certain strict constraints. Thus, it is far from the true trend practical level in the non-ideal controllable situation.As one of the main research targets of the Application Material Foundation of Shanghai-"Research on Self-Organization Safety Surveillance System Based on ARM and RFID" under the project grant number 0512, starting from the key issues which need to be solved in AFR systems, this study plays emphasis on the real-time face detecting and tracking, the face key features localization, the highly effective person face representation, the robust human face recognition classifiers and AFR system design and so on.Face Detection is a key issue which needs to be solved in AFR system design. Many algorithms have been proposed in the few past decades. The algorithm based on the Haar-like Rectangle Feature Cascade Strong Classifiers proposed by Viola in 2001 symbolized the human face detection start to move towards practical. This algorithm detects human faces by extracting the Haar-like features of a human face and training a classifier. It only uses the gradation value without considering the skin color distribution of a human face. Thus it has poor robustness to those face structure-like objects under complex background. In view of this, we propose a real-time face detection algorithm based on skin color model verification and the Haar-like features cascade strong classifier (SCC-HCC) in the third chapter.Face Tracking is an essential stage in those applications, such as video-based human face recognition, video-based monitoring. The CamShift algorithm has strong robustness regarding the object tracking, but it has the drawbacks which Tracking-window must be initialized manually, and the robustness to those skin color-like regions is unsatisfactory. In the third chapter of this thesis, we propose a real-time face tracking algorithm through making improvements to CamShift, named AM-CamShift, which can initialize the track window and implement the multi-objects tracking automatically, and enhance the robustness to skin color-like area.The linear discriminant analysis (LDA), especially the Fisher linear discriminant analysis更多还原