【作者】 柴智；

【导师】 刘正光；

【作者基本信息】 天津大学 ， 检测技术与自动化装置， 2010， 博士

【摘要】 自动人脸识别具有巨大的应用前景,已经成为模式识别、计算机视觉以及信息技术相关学科中活跃的研究领域。过去的几十年中,已经有多种人脸识别方法被提出。本文针对人脸识别中的特征提取和分类问题进行了研究,应用子空间方法,提出了新的人脸识别算法。图像变换是重要的人脸特征提取方法。双树复小波变换可以提取多尺度,多方向的图像特征,同时具有平移不变特性。相对于Gabor小波,双树复小波变换可以有效地保持频域信息。本文提出了基于双树复小波变换和独立成分分析的人脸识别方法。人脸图像的双树复小波变换系数构成特征向量,通过独立成分分析进行子空间投影,采用基于概率推理模型的分类器进行分类。双树复小波变换变换在每个尺度上提取6个固定方向的特征,其中不包括水平和竖直方向。人脸中的主要器官呈现水平和竖直方向特征,提供重要的分类信息。针对双树复小波变换不能直接提取水平和垂直两个方向特征的不足,本文提出了一种结合双树复小波特征和Gabor小波特征的人脸识别新方法。采用0和90的Gabor小波滤波器提取人脸图像水平和竖直方向的特征,与双树复小波特征共同构成特征向量。应用fisherfaces方法进行子空间投影,采用基于欧式距离的分类器实现分类。贝叶斯人脸识别方法分别求取人脸图像的类内和类间差异,在主成分分析子空间中应用高斯概率分布对人脸差异进行建模,通过估计人脸图像差异的后验概率进行分类。由于人脸图像受到多种复杂因素的影响,高斯模型不能对人脸差异进行准确的描述。针对这一问题,本文提出了基于独立成分分析的贝叶斯人脸识别方法。采用独立成分分析计算人脸差异子空间,分别估计每一维度上的概率密度函数,各个维度相乘获得子空间中的高维概率密度函数。本文采用广义高斯模型对人脸差异在独立成分分析子空间中的投影进行建模,提高了概率密度函数估计的准确性和人脸识别方法的效果。更多还原

【Abstract】 Automatic recognition of human faces has been an active research area in the communities of pattern recognition, computer vision and information technology due to its immense application potential. A large numbers of methods have been proposed for face recognition in the last several decades. Feature extraction and classification in face recognition are researched in this dissertation. Novel face recognition methods are proposed by using the subspace methods.Image transform is one of the important face feature extraction methods. The multi-scales and multi-directions image features extracted by dual-tree complex wavelet transform (DT-CWT) are shift invariant. Compared with Gabor wavelet, DT-CWT features preserve more information in frequency domain. A new face recognition method is proposed by adopting the DT-CWT and independent component analysis (ICA). The face feature vectors are constructed by the DT-CWT coefficients of the face images. Subspace analysis is applied to feature vectors by using ICA. Feature vectors in ICA subspace is processed by probabilistic reasoning models for classification.Features on horizontal and vertical directions are not included in the 6 fixed directions features which extracted by DT-CWT on each scale. The horizontal and vertical features provide important information for classification because that the key apparatus on human face are distributed on the two directions. Motivated by the deficiency of DT-CWT, a novel face recognition method is introduced by combining the DT-CWT features and the Gabor wavelet features. The face features on horizontal and vertical orientations are extracted by 0°and 90°Gabor wavelet filters. Face feature vectors are conducted by connecting the 0°and 90°Gabor wavelet features and the features extracted by DT-CWT. Subspace analysis is applied to feature vectors by fisherfaces method. Euclidean distance based classifier is exploited to classify the feature vectors in subspace.Bayesian face recognition calculates the intrapersonal and interpersonal variations of the face images. The face variations are modeled by Gaussian probability distribution in the principal component analysis (PCA) subspace. The posteriori probability of the face variations is estimated for classification. However, Gaussian distribution can not describe the face variations accurately due to the complexity of the influence factors for face image. To overcome this deficiency, an ICA based Bayesian face recognition method is presented. Subspace analysis for face variations is applied to feature vectors by exploiting ICA. One dimension probability distributions are estimated separately in the ICA subspace. High dimension distribution is derived by the multiplication of the mutual independent one dimension distributions. The generalized Gaussian distribution is adopted to model the face variations in ICA subspace. The accuracy of the probability distribution estimation and the performance of the face recognition method are improved by the proposed method.更多还原