【作者】 张志伟；

【导师】 杨瑞霞； 杨帆；

【作者基本信息】 河北工业大学 ， 微电子学与固体电子学， 2007， 博士

【摘要】 近年来在计算机视觉与模式识别领域,人脸识别受到高度关注。作为一种主要身份识别手段,人脸识别在公共安全、视觉监控、数字身份认证、电子商务、多媒体和数字娱乐等领域具有广泛的应用前景。经过近四十年的发展,在人脸识别研究领域已经取得许多成果,已基本实现特定环境下的准确识别,并陆续出现了一些优秀的商用识别软件。尽管如此,人脸识别技术要达到完全实用水平,尚有许多问题有待解决,光照问题就是其中之一。本文针对人脸识别中光照问题进行深入研究。重点研究了可变光照条件下的人脸图像预处理、人脸特征点定位、人脸特征提取以及人脸分类问题。论文主要研究工作如下:(1)研究了可变光照条件下的人脸图像预处理。光照是影响人脸识别系统性能的主要瓶颈之一,由于光照变化导致人脸图像的类内差异要远大于人脸图像的类间差异,为此在光照-反射率模型的基础上提出了一种基于光照参考模型的预处理算法。该算法首先利用多个人的标准光照条件下的人脸图像建立光照参考模型,然后根据参考模型对输入的测试人脸图像进行光照校正,最后利用高斯差分滤波器对校正后的图像进行平滑处理。实验表明该算法可以有效地减少光照对人脸识别的影响,提高人脸识别系统的识别率和光照鲁棒性。(2)研究了可变光照条件下的人脸特征定位。面部特征精确配准是鲁棒实用的人脸识别系统的基本前提,主动形状模型(ASM)和主动表观模型(AAM)是目前解决该问题的主流模型。但它们存在对光照鲁棒性差的缺点,为此提出了一种可变光照条件下的人脸特征定位算法,算法首先利用相位一致性边缘图对人脸图像的瞳孔位置进行粗定位,根据瞳孔的位置进行模型初始化,然后利用光照不敏感的特征进行统计外观建模,对人脸图像进行粗定位;最后在粗定位的基础上,将人脸图像分为多个子区域,在各个子区域上应用AAM算法进行精确定位。实验表明该算法能够比较精确的定位面部关键特征,对光照变化具有较好的鲁棒性。(3)研究了可变光照条件下的人脸特征提取。特征提取是人脸识别中关键的一步,必须对光照、姿态、表情、化妆、年龄等变换具有较强的鲁棒性。为此先对子空间特征提取算法——局部保持投影(LPP)进行改进,提出了一种有监督LPP算法,即对每个LPP基向量进行线性判别分析,选择主要反映类间差异的基图像来构造新的子空间。然后将监督LPP算法扩展到二维,提出了一种二维有监督LPP算法。实验结果表明在可变光照条件下,这两种改进算法的识别率都高于传统的子空间算法,说明这两种算法对光照具有一定的鲁棒性。此外,基于Gabor小波具有良好的局部特征和方向选择性,且对光照与姿态具有较强的鲁棒性,提出了一种融合Gabor相位和幅值信息的子空间特征提取算法。该算法同时具有Gabor方法和子空间方法的优点,对光照变化具有很强的鲁棒性。(4)研究了可变光照条件下的人脸分类问题。分类器的设计是人脸识别中的最后环节。常见的分类器有最近特征线分类器(NFL)和支持向量机(SVM)等,但它们均基于欧氏距离来进行相似度量,所以不能用于度量光照等非线性因素。为此对最近邻分类器和最近特征线分类器进行改进,即利用能刻画非线性流形特征的测地线距离代替传统的欧式距离,最终得到一种最近测地线分类器(NGC)和一种基于测地线距离的最近特征线分类器(NFLG)。实验结果表明,这两种改进的分类器能够有效地提高识别率。更多还原

【Abstract】 Face recognition has attracted considerable interests in the recent years within the community of computer vision and pattern recognition. As one of the most successful branches of personal identification, it has great potential applications in public security, visual surveillance, digital personal identification, electronic commerce, multimedia, and digital entertainment, etc. The face recognition has developed rapidly over the past 40 years. Now under the controlled conditions, face recognition systems have achieved good results, and some versatile commercial recognition software have been appeared. However, face recognition technologies are currently far from mature. A great number of challenges are still leaved to resolve before one can implement a robust and practical face recognition application. Among these challenges, the illumination circumstance is one of the most difficult.Our work is focusing on the effect of illumination on the face recognition. The emphases of the work are the image preprocessing, feature point location, feature extraction and the classification under the varying illumination circumstance. The work and the innovation in this dissertation can be summarized as following.(1) Facial image preprocessing under varying illumination was studied.The illumination is one of the bottlenecks that affect recognition performance. In most cases, the difference between two images caused by illumination is greater than that caused by individual difference. Motivated by the illumination-reflectance model, a preprocessing algorithm based on the illumination-reference is proposed. Firstly, the algorithm constructs an illumination reference model utilizing a set of face images of different persons under normal illumination conditions. Following that, the testing face images are adjusted according to the illumination reference model. Lastly, the difference between Gaussian filters is used to smooth the boundary of the images that has been adjusted. Experimental results exhibits that the proposed algorithm could reduce the effect of the illumination. Meanwhile, the accuracy and robustness of the recognition system are improved.(2) Facial feature alignment under varying illumination was studied.Accurate facial feature alignment is the prerequisite of a face recognition system. Currently, the Active Shape Model (ASM) and Active Appearance Model (AAM) are the main models for this problem. However, two models mentioned above are sensitive to the illumination variation. To fight with the disadvantages of two models, an improved AAM under varying illumination is proposed. Firstly, the eyes are located by using the phase congruency binary edge image, which is used to initialize the model. Secondly, the model is constructed by utilizing features that are illumination robustness. After that, face is aligned coarsely. Finally, the facial images are segmented into some sub-regions, thus the improved AAM is used to get fine location in every sub-region. Encouragingly, experiment results have illustrated the better performance and illumination robustness of the proposed model.(3) Facial feature extraction under varying illumination was studied.Feature extraction is a key step to face recognition. The extracted feature should be robust to illumination, poses, expression, and age, etc. We generalize the conventional LPP algorithm to yield a new supervised algorithm. We name it as supervised LPP, which can deal with linear discriminant analysis in the features derived from NMF. In this way, a discriminant feature subspace having the maximum intersubject variation and the minimum intrasubject variation is established. Following that, the supervised LPP is extended to the two dimensions, thus a new two dimension supervised LPP is introduced. The results in numerical experiments show that the two proposed algorithms have higher recognition rates than the traditional subspace algorithms under varying illumination, which show they are illumination robustness.The Gabor wavelet has good characteristics of spatial location and orientation selection. Based on these observations, a Gabor subspace feature extraction algorithm is introduced, which attempts to fuse the Gabor phase and magnitude information. The proposed algorithm blend the merits of Gabor wavelet-based and subspace-based algorithm. In addition, it is illumination robustness.(4) Classification problem under varying illumination was studied.The classifier design is the last step of the face recognition. The Nearest Neighbor (NN) and Support Vector Machine (SVM) are traditional classifiers, which uses the Euclidean distances in the similarity measurement between different features. However, the Euclidean distances can’t be used to measure the nonlinear manifold of face images when the illumination conditions change frequently. Replacing Euclidean distances by the geodesic distances, a novel nearest geodesic distances classifier and a nearest feature Line classifier using geodesic distances are presented. It has been shown from the experimental results that the two novel classifiers have higher recognition rates than the traditional classifiers.更多还原