【作者】 李外云；

【导师】 刘锦高；

【作者基本信息】 华东师范大学 ， 无线电物理， 2008， 博士

【摘要】 嵌入式人脸识别系统建立在嵌入式操作系统和嵌入式硬件系统平台之上,具有起点高、概念新、实用性强等特点。它涉及嵌入式硬件设计、嵌入式操作系统应用开发、人脸识别算法等领域的研究;嵌入式人脸识别系统携带方便、安装快捷、机动性强,可广泛应用于各类门禁系统、户外机动布控的实时监测等特殊场合,因此对嵌入式人脸识别的研究工作具有突出的理论意义和广泛的应用前景。本文是上海市经委创新研究项目《射频识别RFID系统-自动识别和记录人群的身份》(编号:04-11-2)与上海市科委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。论文从构建自动人脸识别系统所需解决的若干关键问题入手,重点探讨了基于嵌入式ARM微处理器的实时人脸检测、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的研究。论文的主要工作和创新点表现在以下方面:1实现了结合肤色校验的Haar特征级联分类器嵌入式实时人脸检测,提出了基于人脸约束的人眼Haar特征RSVM级联分类器人眼检测算法和基于遮罩掩磨与椭圆拟合的瞳孔定位算法。复杂背景中的人脸检测是自动人脸识别系统首先要解决的关键问题,通过对基于肤色模型和基于Haar特征级联强分类器的人脸检测算法的分析研究,综合两个算法的优点,提出了基于肤色模型校验和Haar特征级联强分类器的嵌入式实时人脸检测算法。实验结果表明,该算法不仅解决了复杂背景中的类肤色和类人脸结构问题,而且具有较高的检测率和较快的检测速度,同时对光照、尺度等变化条件下的人脸检测也具有较强的鲁棒性。人眼检测与瞳孔定位在人脸归一化和有效人脸特征抽取等方面起着非常重要的作用,为了快速检测人眼并精确定位人眼瞳孔中心,论文提出了基于人脸约束的人眼Haar特征RSVM级联分类器人眼检测算法和基于遮罩掩磨与椭圆拟合的瞳孔定位算法,首先利用人眼检测分类器在人脸区域内完成对人眼位置的检测,然后通过对检测到的人眼进行遮罩掩磨、简单图像形态学变换及椭圆拟合实现瞳孔中心的精确定位。测试结果表明该算法只需几百毫秒便能完成人眼检测与瞳孔中心定位整个过程,在保证检测速度较快的同时,还能确保较高的定位精度。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算法及相应的人脸识别方法人脸识别中的小样本问题使线性判别分析算法的类内散布矩阵发生严重退化,导致问题无法求解。本文在人脸识别小样本问题的基础上,通过调整Fisher判别准则,利用类间散布矩阵的补空间巧妙地避开类内散布矩阵的求逆运算,通过训练集每类样本的样本数信息自适应改变调整参数,实现了自适应线性判别分析算法,实验结果表明,该算法能有效解决人脸识别中的小样本问题。3提出了基于有效人脸区域的Gabor特征抽取算法,有效地解决了Gabor特征抽取维数过高的问题。Gabor小波对图像的光照、尺度变化具有较强鲁棒性,是一种良好的人脸特征表征方法。但维数过高的Gabor特征造成应用系统的维数灾难,为解决Gabor特征的维数灾难问题,论文第四章提出了基于有效人脸区域的Gabor特征抽取算法,该算法不仅有效地降低了人脸特征向量维数,,缩小了人脸特征库的规模,同时降低了核心算法的时间和空间复杂度,而且具有与传统Gabor特征抽取算法同样的鲁棒性。4结合有效人脸区域的Gabor特征抽取、自适应线性判别分析算法和基于支持向量机分类策略,提出并实现了基于支持向量机的嵌入式人脸识别和嵌入式人像比对系统支持向量机通过引入核技巧对训练样本进行学习构造最小化错分风险的最优分类超平面,不仅具有强大的非线性和高维处理能力,而且具有更强的泛化能力。本文研究了支持向量机的多类分类策略和训练方法,并结合论文中提出的基于有效人脸区域的Gabor特征提取算法、自适应线性判别分析算法,首次在基于Windows CE操作系统的嵌入式ARM平台中实现了具有较强鲁棒性的嵌入式自动人脸识别系统和嵌入式人像比对系统。5提出并初步实现了基于客户机/服务器结构无线网络模型的远距离人脸识别方案为解决嵌入式人脸识别系统在海量人脸库中进行识别的难题,论文提出并初步实现了基于客户机/服务器结构无线网络模型的嵌入式远距离人脸识别方案。客户机(嵌入式平台)完成对人脸图像的检测、归一化处理和人脸特征提取,然后通过无线网络将提取后的人脸特征数据传输到服务器端,由服务器在海量人脸库中完成人脸识别,并将识别后的结果通过无线网络传输到客户机显示输出,从而实现基于客户机/服务器无线网络模型的嵌入式远距离人脸识别方案。6结合我们开发的基于ARM的嵌入式自动人脸识别系统和嵌入式人像比对系统,从系统设计的角度探讨了在嵌入式系统中进行人脸识别应用设计的思路及应该注意的问题虽然嵌入式人脸识别系统的性能很大程度上取决于高效的人脸特征描述和鲁棒的人脸识别核心算法。但是,嵌入式系统的设计思想对嵌入式人脸识别系统的性能影响同样值得重视。本文第六章重点阐述了嵌入式自动人脸识别应用系统的设计思路,并结合我们自主开发的嵌入式自动人脸识别系统和嵌入式人像比对系统从系统设计的角度探讨了嵌入式人脸识别应用系统设计中应该注意的关键技术问题。结合本文提出的算法我们在PC上完成对人脸识别分类器的训练,然后在嵌入式ARM开发平台上实现了嵌入式自动人脸识别、嵌入式人像比对两个便携式人员身份认证系统,经测试运行效果良好。所提出的人脸识别算法不仅具有一定的理论参考价值,而且对于嵌入式系统应用开发、AFR应用系统开发也具有一定的借鉴意义。更多还原

【Abstract】 Embedded human face recognition is built on the embedded operating system and embedded hardware platform,which involves embedded hardware design,embedded operating system application development,human face recognition algorithms and so on.It is a high starting point,the new concept,practicality AFR.As one of easy to carry,quick installation and mobile AFR,it can be widely applied to different kinds of occasions such as access control systems,outdoor mobile real-time monitoring and other special occasions,so research on embedded face recognition has a strong theoretical significance and wide application.As one of the main research target innovative research projects of Shanghai Municipal Economic Commission-"Radio Frequency Identification RFID system-automatically identify and record the identity of the crowd" (No.04-11-2)and AM Fund project of Shanghai Science and Technology Commission-"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 embedded AFR systems,this study plays emphasis on the real-time face detection,the face key features location,the highly effective person face representation,the robust human face recognition classifiers and AFR system design and so onAutomatic human face detection under the Complex background is the first key issues need to be resolve in the AFR systems,through study the human face detection algorithm base on human skin color model and Haar-like rectangle feature cascade strong classifiers,we found that the face detection algorithm base on human skin color model only use the skin color information without considering the gradation value,and Haar-like rectangle feature cascsde strong classifiers on the country,it only use the gradation value without considering the human skin color information.Thus they have poor robustness to those color-like and face structure-like object 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 in the second chapter.The results show that the algorithm not only solve the color-like and face structure-like object problems under complex background,but also has high detection rate and faster detection rate,and it is robust to light,scales changes under complex background.The human eye detection and pupil location play a very important role on human face normalization and effective human face feature extraction. In order to rapid detect the human eye and precise positioning the human eye pupil center;we propose human eye detection algorithms base on Haar-like features RSVM cascade classifier,and the pupil location algorithms base on the mask and elliptical fitting,the experimental results show that it only take a few hundred milliseconds to complete the human eye detection and pupil centre location the whole process by use the new algorithms,it has fast detection rate and high location accuracyThe same sample size problem in human face recognition will degrade spread matrix in linear discriminant analysis algorithms,and will lead to the problem can not be solved.To solve this problem,we proposes the adaptive linear discriminant analysis algorithm through adjusting the Fisher criterion and making the Improvement to the Fredman thought,Using the complement space of between-class scatter Matrix the algorithm avoids the inverse operation of within-class scatter matrix and adaptively changes the parameter according to the sample information of each class. The experimental result shows that the adaptive linear discriminant analysis algorithm can resolve the SSS problem of FR effectivelyGabor wavelet is robust to image light,scale changes,and it is a good facial feature characterization.But the dimension of excessive Gabor characteristics will lead to dimension disaster of the application system, in order to solve this problem,we proposes Gabor feature extraction algorithms base effective human face region.This algorithms not only reduces the human face feature vector dimension effectively,make small the human face library scale,while reducing the core algorithms of time and space Complexity.And it has same robustness with the traditional Gabor feature extraction algorithm.The support vector machine(SVM),anewmethod for data mining in recent years,has its unique superiority in many fields,such as pattern recognition and nonlinear programming and so on.In this paper,we study multiple classifier strategy and training method of the SVM,and combining with Gabor feature extraction algorithms,adaptive linear discriminant analysis algorithms,develop a strong robustness embedded automatic face recognition system base Windows CE operating systems in ARM platform.To resolve the difficult problem of embedded automatic face recognition and identify in massive face library,we propose a preliminary remote Face Recognition programme which based on client/server architecture wireless network modelThe human face detection,pre-process,normalization and the face feature extraction is completed in the client terminal(embedded hardware platform),and then the client transmits the human face feature data to the server through wireless network.After completing the face recognition and identify in massive face library,the server transmit the result through the wireless network to the client for display.The final performance of embedded FR application system is decided in the very great degree on the highly effective face representation and the robust FR core-algorithm,but the system design strategies are also worthy to be paid attention to.In the ChapterⅥ,this paper narrated the embedded AFR system design mentality in detail,and discusses the key technical issues in embedded AFR system design with two FR prototype systems demonstrated from the viewpoint of system design.Finally,based on our proposed embedded FR core-algorithms,we realized two embedded FR systems,embedded Intelligent Video Surveillance system and facial image matching system.The systems show good performances in demonstrations and validations.The proposed algorithms can not only contribute to AFR theory,but also have reference values for embedded AFR application system design.更多还原