【作者】 邹国锋；

【导师】 王科俊；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2013， 博士

【摘要】 人脸识别一直是模式识别和机器视觉领域的热点研究问题。人脸作为一种独特的生物特征，具有直接性、唯一性、便捷性等特点，但是由于人脸的可塑多变和在成像过程中受诸多因素的影响，又使得人脸的机器自动识别这一课题极具挑战性。特别是非受控环境下人脸的多姿态变化问题、摄像机更换以及人体运动等因素带来的人脸图像退化（低分辨率和模糊）问题，这些亟待解决的关键问题已成为人脸识别技术在视频监控环境下得到广泛应用的瓶颈之一。本文主要针对人脸识别中的多姿态变化和退化两个关键问题开展了相关的研究。多姿态人脸识别主要涉及两大类型：（1）可通过几何归一化得到校正的平面内旋转人脸识别；（2）无法进行几何校正的俯仰变化和左右摇摆变化的人脸识别。针对这两大类型的多姿态人脸，本文分别提出了有效的人脸识别方法。针对退化人脸识别问题，本文主要从距离度量的角度出发，提出了一种非线性的耦合度量策略，它能够直接用于不同分辨率图像和模糊图像的特征提取与分类识别。本文的主要研究内容总结如下：首先，针对人脸在同一个平面内的旋转变化问题，提出具体的解决方案。首先，提出一种融合Adaboost算法和分块积分投影的眼睛定位方法，实现了对倾斜人脸中眼睛的高精度定位。该方法采用Adaboost人脸和眼睛分类器粗估计出人眼区域，然后，在人眼区域中通过分块积分投影法准确定位双眼位置。确定双眼位置后，给出以图像几何中心为旋转基准点进行图像旋转校正时旋转角度的计算方法，实现了以图像几何中心为旋转基准点对倾斜人脸图像进行的平面校正。这对于完善以图像中心为基准的图像校正方法有一定的意义。最后，提出一种基于子模式的Gabor特征融合的人脸识别方法，有效提取了校正后人脸图像中的特征信息用于分类识别或相似度计算。其次，针对人脸多姿态变化中无法通过几何校正来克服的俯仰变化和左右摇摆变化问题，提出一种全新的人脸表征方法，即人脸能量图（Face energy image，FEI），并分别定义了广义和狭义人脸能量图，通过理论证明说明了人脸能量图所具有的多个优点；针对人脸能量图模糊的问题，采用改进的Retinex图像增强方法对其进行了增强预处理；最后，针对人脸能量图存在特征冗余的问题，提出一种基于最大分离度差的有监督局部保持投影特征提取方法，能够有效提取人脸能量图中所蕴含的非线性流形信息和分类信息用于识别。然后，在本文提出的人脸能量图（人脸均值能量图）基础上，进一步提出另外一种新的能量图，即人脸方差能量图（Face variance energy image，FVEI）。人脸均值和方差能量图分别从均值和方差两个角度对多姿态人脸图像进行了描述。由于这两种能量图在人脸识别过程中具有不同的分类作用，所以，以这两种不同含义的特征为基础，结合特征级融合的策略，本文提出一种有效融合均值和方差能量图的多姿态人脸识别方法，能够更好的解决姿态变化给人脸识别带来的困难。最后，本文将人脸识别中出现的退化人脸识别问题看作不同数据集合元素间的度量问题，针对不同数据集合元素间的度量问题，提出一种新的基于有监督局部保持投影的耦合度量学习方法，并在此基础上与核技术相结合，将该方法推广到了非线性情况下，提出基于有监督局部保持投影的核耦合度量学习方法。分别将提出的这两种耦合度量学习方法应用到退化人脸识别中，取得了较好的识别效果。更多还原

【Abstract】 Face Recognition has been a hot issue in the field of pattern recognition and machinevision. As a unique biometric feature, face has the characteristics of directness, uniqueness,convenience etc. But due to the plasticity, variability and influence of many factors in imagingprocess, the automatic face recognition becomes a challenging task. Especially in videomonitoring environment, the problems that the face pose change in uncontrolled environmentand the degradation of face image (low resolution and fuzzy) because of camera replacementor human motion have become one of the bottlenecks of face recognition.In this paper, the relevant research on multi-pose changes and face degradation has beencarried out. Multi-pose face recognition includes two types:(1) The plane rotation facerecognition which could be obtained by the normalization;(2) The face recognition in thecase of pitch change and vacillating change which can not be corrected by the normalization.To solve the problem of these two types of multi-pose face recognition, effective facerecognition methods were proposed respectively. For the degradation problem of facerecognition, we proposed the idea of nonlinear coupled metric. The nonlinear coupled metriccan be directly used for feature extraction and classification of different resolution images andblurred images. The main contents of this paper are as follows:Firstly, aiming at the problem of face plane rotation, specific solution was proposed.Firstly, an eyes location method combining the AdaBoost algorithm with block integralprojection was proposed. This method has realized the high precise location of eyes ininclined face. The eye region could be roughly estimated based on the AdaBoost faceclassifier and eye classifier, and then accurate positions of two eyes were located in the eyearea through block integral projection. After locating the eyes position, the calculation methodof the rotation angle was given when the geometric center of image was as the rotationreference point. We realized the face plane rotation correction. The proposed method hascertain significance for perfecting the image correction theory. Finally, a face recognitionmethod based on sub-pattern Gabor features fusion was proposed, and it can effectivelyextract the feature of corrected face image which were used to the classification and similaritycalculating.Secondly, aiming at the problem of pitch change and vacillating change in multi-posechanges which could not be overcome through geometric correction, we proposed a novelface representation method, which is face energy image (FEI). Then the generalized face energy image and narrow face energy image were defined. The advantages of face energyimage have been verified through theoretical proof. For the fuzzy problem of face energyimage, we carried out image preprocessing adopting improved Retinex image enhancementmethod. Finally, to solve the problem of feature redundancy, a supervised locality preservingprojection feature extraction method based on the maximum separation degree difference wasproposed, which could be used to extract the nonlinear manifold information and classinformation contained in face energy image.Then, on the basis of face energy image (face mean energy image), another new energyimage was proposed, which is face variance energy image (FVEI). Face mean energy imageand variance energy image describe the multi-pose face image from two different angles:mean and variance respectively. Because these two kinds of energy image have differentclassification effects, on the basis of these features, combining with the feature level fusionstrategy, we proposed an effective multi-pose face recognition method integrating mean andvariance energy image. This method can solve the difficulties of multi-pose face recognitionbetter.Finally, in this paper, the degradation problem of face recognition was regarded asmeasurement problem among elements of different data collection. In order to solve thismeasurement problem, a new coupled metric learning method based on supervised localitypreserving projection was proposed. Combined with kernel technique, the method wasextended to nonlinear situation, futher more the kernel coupled metric learning method basedon supervised locality preserving projection was proposed. Finally, these two proposedcoupled metric learning methods were respectively applied to deal with the degradationproblem of face recognition, and better recognition effect was obtained.更多还原