【作者】 吴翔；

【导师】 刘锦高；

【作者基本信息】 华东师范大学 ， 无线电物理， 2009， 博士

【摘要】 人脸识别是基于生物特征的身份认证技术中最为活跃最具挑战性的领域之一,有着很大的发展潜力。它试图通过计算机分析比较人脸图像并从中提取有效特征信息进行身份鉴别。由于利用人脸特征进行身份验证是最自然直接的手段,易于为用户所接受。特别是在非接触环境和不惊动被检测人的情况下,人脸识别技术的优越性远远超过已有的虹膜、指纹等检测方法,因而成为当前模式识别和人工智能领域的一个研究热点。目前的人脸识别方法基本上都集中于二维图像方面,这是由现在照相和摄影技术先天的性能所决定的。但是人脸本身是三维空间的实体,因此在三维转二维的过程中,丢失了很多有用的立体信息,这会造成识别误判。关于传统的基于二维平面内的人脸识别方法,国内的研究有很多,这些算法在二维图像识别的领域,确实已经达到了非常好的识别效果。可是丢失人脸三维景深信息会导致二维人脸识别方法很难摆脱人脸塑性变形(如表情等)的不确定性、人脸模式的多样性(如胡须、发型、眼镜、化妆等)、图像获取过程中的不确定性(如光照的强度、光源方向等)的影响。在复杂环境下系统会表现出识别率下降,识别效率降低。这些极大限制了人脸识别的应用。所以到目前为止,建立一个鲁棒的人脸识别系统仍然是一个很困难的问题。为了解决这个问题,一种行之有效的方法就是采用三维人脸识别的方法解决二维人脸识别中面临的光照和姿态的瓶颈问题,减少表情的影响。三维人脸模型具有比二维人脸图像更丰富的信息。3D数据真实反映了物体在三维空间中的形状,表征了对象的实际尺寸,具有对姿态和光照不敏感的特性。三维人脸识别中最重要的部分就是建立对应的三维人脸模型,而如何获得形状模型和精确的纹理影射是构建模型的关键。目前在人脸重建领域还没有一种既采样简单又计算精确的获得人脸三维信息的方法。在人脸识别领域,随着3D数据获得技术如激光扫描、双目和多目视觉等的进步,3D数据研究受到更多的关注。如何能够简单高效的获得和使用三维信息,成为了研究的焦点。本文基于DLP立体显示系统的需要,对其三维图像源的获取、合成、投影及光路控制等进行了深入研究:通过对双目摄像头采集的立体动态视频序列的研究,给出了如何存储和表征人脸信息并生成人脸彩色图和深度图的解决方案;利用上述肤色和深度信息,可圈定面部区域,并获得人脸特征点的深度信息及点阵之间的拓扑结构;尝试通过2D和3D相结合的方法提取人脸特征点,利用两幅不同角度的二维图像重建三维人脸模型。通过对重建的三维人脸模型的旋转和平移变换,并利用投影矩阵亦可得到多姿态的二维人脸图像,使得其投影视图更加接近人脸库中的原型,弱化了人脸拍摄时的位置影响。最后通过高精度有效的识别算法匹配完成对人脸的识别过程。通过更全面的信息,可以较好的降低识别过程中的误识率和虚警率。同时由于三维人脸模型具备光照无关性和姿态无关性的特点,能够正确反映出人脸的基本特性,从而形成相对稳定的人脸特征表述。因此基于本文提出的三维人脸模型的识别方法可以很好的解决目前在这一领域存在的研究瓶颈。本文是上海市2007年科技攻关重点项目(《基于单片DMD的立体显示系统》编号:075115002)的主要研究内容之一,并受到华东师范大学优秀博士研究生培养基金(20080050)支持。从构建基于立体显示的三维人脸识别系统需要解决的若干关键问题入手,重点讨论了双目视觉计算景深的原理、人脸检测与跟踪、立体图像对中点与点的匹配、面部关键特征定位、三维的人脸特征描述、鲁棒的人脸分类器以及整个系统软硬件设计等问题。从这几个方面入手,深入研究了基于图像和模型的人脸模型重建方法,人脸标志点定位及人脸配准。本文的创新点有:1.提出基于双目立体测量系统提取三维人脸信息的方法,分析了双目立体测量系统的基本原理,深入研究双目立体测量系统在测量中物像空间点位的精确关系,探讨了物像空间点位坐标关系标定、测量范围扩展等关键技术,并推出精度问题。2.提出利用相关系数、欧式距离和ICP结合的方法,并通过形状特征定位五官分布位置作为几何约束条件,实现双目视觉中的人脸立体图像对内对应点的匹配。3.提出基于人脸结构的立体匹配算法,配准人脸特征模型,通过对模型相似度的度量实现人脸识别。改进2D人脸识别中的PCA、LDA等经典算法到3D人脸识别中,结合ICP和Hausdorff距离识别算法,完成识别过程并验证识别效果和精度。4.提出硬件和软件设计实现方案。规划一个完善的系统,包括电源、存储器、处理器、控制器、通信接口及人机交互接口。综合考虑电磁兼容性、系统稳定性等。代码流程、接口及部分模块的详细源码也在本文中详细给出。更多还原

【Abstract】 Face recognition is one of the most active and challenging area of all authentication technologies based on biological characteristics,which will have great potential for development. It extracts valuable information from different face pictures by computer for identification.This technology is widely applied in information security,video surveillance,target tracking and other areas.But limited by camera and photographic technology,the current face recognition methods have focused on two-dimensional images.Essentially,Human face is three-dimensional object. When changing the 3D objects to 2D pictures,a lot of useful ID information is lost.So there are unavoidable defects in 2D AFR,such as the distortion for different expressions,the diversity of human face patterns(e.g.beard,hair,glasses,makeup,etc.)and the uncertainty for complex environment(e.g.light intensity and direction of light source,etc.).To solve this problem,an effective method is using 3D AFR algorithm instead of 2D AFR. The three-dimensional model contains more information than two-dimensional facial image.3D data gives an accurate reflection of the objects,such as shape and size,and it is not sensitive to the posture or illumination.The most important part in 3D recognition is to establish the three-dimensional model,and how to get the shape model and accurate texture projection are crucial.Currently there is no easy way to obtain the three-dimensional information accurately in the field of facial reconstruction.Overall,more attention should be paid to how to get three-dimensional information more simply and efficiently in 3D technology research.In this paper we try to realize the face recognition using 3D data which are from the video sequence obtained by two cameras.We develop a novel way to get,depict and store 3D human face information,generate both color maps and depth maps of the face.Face region is tracked, based on this region outline and depth of feature points is extracted to form the topological structure.We propose novel 2D+3D method to get the facial feature points,and rebuild three-dimensional face model by two 2D pictures from binocular cameras.Through translation and rotation of the three-dimensional model,we can get two-dimensional face picture of all attitude by the projection matrix.That means the direction of the face can be adjusted and the influence of the position can be weakened,its projection will be closer to the prototype in face database.Finally an effective high-precision matching algorithm is given to complete recognition.With full-scale information,FAR(False Accept Rate)and FRR(False Reject Rate)have decreased a lot.Meanwhile,for the three-dimensional model has nothing to do with illumination and posture,the system can reflect the basic characteristics of human face correctly,the 3D AFR algorithm this paper post is stable and robust.The research based on three-dimensional face model in this article can be a very good solution to bottlenecks existing in current AFR field.This paper is one of the main contents of the key scientific and technological project of Shanghai in 2007—“Nakedness-Eye Stereoscopic Display Based on Single Chip DMD”under the project grant number 075115002.It’s also supported by outstanding PH.D educating Fund of East China Normal University under the grant number 20080050.We start from the key issues of 3D face recognition system,to play emphasis on the binocular vision principle,face detection and tracking,the point matching in 3D image,face features location,3D face representation,the robust human face classifiers and the hardware design for whole system and so on.This thesis’s innovative points include:1.We analyze the principle of binocular 3D measurement system,the relationship of the objects’ depth and disparity is researched.Then we propose a method to extract 3D facial information in stereoscopic image pair and extend to the topic about accuracy.2.A method about related coefficient,Euclidean distance and ICP was given to resolve the difficulty in stereo matching of two binocular face imge pair.3.We propose the 3D matching algorithm based on human face structure.ICP and Hausdorff distance are used,along with improved PCA,LDA,Gabor and other classical algorithm in 2D face recognition.4.We propose the realization in both hardware and software design.The complete system includes power supply,memory,processors,controllers,communication interface, interactive interface and EMC analysis.The flowchart and some codes are also given here.更多还原