【作者】 付晓峰；

【导师】 韦巍；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 人们的工作和生活越来越多地和计算机联系在一起,使得人类与计算机之间的关系越来越紧密。另外,各种各样的机器人也相继问世并且会越来越多地出现在我们的周围。人们渴望自然和谐的人机交互——计算机首先要识别主人的身份,然后判别主人的表情以做出相应的动作。因此,本文以快速准确的人机交互为目标,旨在研究如何提高基于人脸的身份识别与表情识别的识别率及识别速度。论文的主要研究工作与成果包括以下几个方面的内容:1)提出一种基于局部二元模式直方图映射(LBPHP)的快速人脸识别方法。此方法将局部二元模式直方图映射到保局投影(LPP)空间获得低维的LBPHP特征,在此低维特征空间判别新样本大大提高了识别速度,由于LPP强大的鉴别特性,此方法的识别率也很高,相比于传统的基于局部二元模式(LBP)的人脸识别方法,此方法不仅识别速度更快、识别率更高,尤其在大型人脸库上凸显其优势,适于此类人脸库上的实际应用如身份识别等。2)在表情特征提取方面,传统LBP算子存在不足:产生的直方图维数过长、鉴别力不高、对噪声反应敏感.针对此问题,提出中心化二元模式(CBP)算子。CBP算子相对于LBP算子具有三大优势:(1)CBP算子通过比较“近邻点对”之间的差异捕捉到梯度信息,不仅增强其鉴别能力而且大大降低特征维数。(2)CBP算子充分考虑中心像素点的作用并给它分配最高权重,此举大大提高其鉴别能力。(3)从图像中提取到的CBP特征在有噪声情形下更加鲁棒、更加稳定。此外,为提高识别结果,首次将中心最近邻分类器引入表情识别中,它的分类效能优于目前常用的最近邻分类器。3)为进一步提高人脸表情识别率,对中心化二元模式(CBP)做拓展:(1)将梯度信息融入CBP;(2)提出多尺度CBP(简称MCBP);(3)为增强算法对表情图像中细小变形的鲁棒性,首次引入图像欧式距离(IMED)并将其嵌入MCBP方法.嵌有IMED的MCBP(简称MCBP-IMED)方法提取出的特征具有优点:维数大大降低、很强的鉴别能力、对噪声不敏感、不易受细小变形的干扰。4)提出结合了CBP与Gabor变换的中心化Gabor(简称CGBP)直方图,并将梯度信息融入其中。为更好地反映表情流形的内在结构,提出融合了局部方法和有监督方法思想的有监督Laplacianfaces(简称SLAP)。另外,注意到人脸面部表情与人们情感表达的密切关系,提出一种融合了连续性与离散性的表情空间模型。基于此模型,将SLAP应用于融入梯度信息的CGBP直方图进行表情识别及表情成分分析。更多还原

【Abstract】 Human’s life and work is related increasingly to computer,so the relationship between human and computer becomes closer and closer.Furthermore,many kinds of robots come out and will appear increasingly around us.Human long for the natural and harmonious human-computer interaction(HCI),i.e.,computers firstly distinguish the mastership and then act according to recognizing host’s expression.Hence,the thesis sets the fast and accurate HCI as the goal and aims at improving the accuracy and speed of the face recognition and expression recognition.The major research works and contributions of the thesis include:1) It proposed a fast face recognition approach called local binary pattern histogram projection(LBPHP).The method projects local binary pattern(LBP) histogram onto locality preserving projection(LPP) space and obtains the low dimensional LBPHP feature.It is fast to recognize new sample in the low dimensional space.The method has good accuracy in the light of LPP’s powerful discriminative property.LBPHP method is superior to conventional method based on LBP not only on recognition speed but also on accuracy.It is prominent especially on large-scale face database and suitable for practical application e.g.,identity authentication.2) In the aspect of expression feature extraction,the conventional LBP operators have several disadvantages such as rather long histograms produced by them,lower discrimination and sensitivity to noise.Aimed at the problems,it proposed the centralized binary pattern(CBP) operator.CBP operator has several advantages:(1) CBP operator captures the gradient information by comparing pairs of neighbors, which not only improves its discrimination but also reduces significantly the feature’s dimensionality.(2) Its discrimination is strengthened due to considering the center pixel point and giving it the largest weight.(3) The CBP feature extracted from image is more robust and more stable in the noisy situation.Moreover,for the purpose of increasing recognition accuracy,it introduced the center-based nearest neighbor classifier into expression recognition.This kind of classifier is superior to the nearest neighbor classifer.3) In order to increase the accuracy of expression recognition,it expanded CBP from the following aspects:(1) The gradient information was integrated into CBP.(2) Multi-scale CBP(MCBP) was proposed.(3) For the purpose of improving the robustness to small deformation of expressional images,it introduced image Euclidean distance(IMED) and embedded it in MCBP.MCBP-IMED is short for the approach of MCBP embedded with IMED.The feature extracted by using this method has advantages:lower dimension,very powerful discrimination,insensitivity to noise, robustness to small deformation.4) It proposed the centralized Gabor binary pattern(CGBP) histogram,which combined CBP and Gabor transform,and integrated the gradient information into CGBP histogram.In order to reflect accurately the underlying structure of expressional manifold,it put forward the supervised Laplacianfaces(SLAP) which combined the ideas of local method and supervised method.Furthermore,because of the close relationship between facial expression and feeling,it proposed the expressional space model which combined the continuity and scatter of expressional space.Baesd on the model,SLAP was applied to CGBP histogram with gradient information to recognize expression and analyse expression components.更多还原