【作者】 刘中华；

【导师】 金忠；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2011， 博士

【摘要】 受公共安全、金融安全以及人机交互等领域大量潜在的需求所驱动,生物特征识别已经成为模式识别和人工智能领域的一个研究热点。尤其人脸识别由于其自然、直观、非接触、安全等特点而倍受关注,成为最具发展潜力的生物特征识别技术之一。经过近几十年的发展,人脸识别领域积累了丰富的理论和大量算法,初步解决了可控条件下的人脸识别难题。然而,在非配合和非控制条件下的人脸识别依然是一个非常具有挑战性的课题。影响识别性能的非控制因素有很多,例如：姿态、光照以及表情等变化,其中光照变化对人脸识别的影响尤其明显。本文主要针对人脸识别中的光照变化问题进行了深入细致的研究。重点研究了光照变化条件下的人脸图像预处理、特征提取等问题。论文主要研究工作如下：(1)对光照子空间和商图像方法进行了研究,提出了一种可变光照下的人脸识别方法。基于三维光照子空间模型的商图像方法过于简单,不能很好地对极端光照条件进行处理,并且不能处理人脸自身阴影。本文通过对商图像和光照子空间的研究,提出了基于九维光照子空间的改进后商图像方法。该方法首先合成图像库中每一个人的九幅基图像,这些基图像可以表示该对象不同光照条件下的人脸图像；其次,利用光照比图像方法合成图像库中每一对象与待识别图像相同光照条件下的一幅虚拟人脸图像；最后,用这些新合成的虚拟人脸图像来完成不同光照条件下的图像识别。(2)研究了光照预处理算法和基于Gabor小波的特征提取算法,提出了一种自适应的Gabor图像特征抽取和权重选择的光照不变人脸识别方法。为了消除光照变化对人脸识别的影响,通过对光照预处理算法的研究,本文首先提出了一种改进的局部对照增强算法。其次,Gabor小波具有较好的方向选择性和空间局部性,它可以捕获图像在不同方向、不同频率下的边缘以及局部显著特征,且对光照具有较强的鲁棒性。因此,Gabor小波被广泛应用于人脸图像识别,以提取鲁棒的人脸特征。然而,Gabor特征的过高维数需要比较大的存储空间并且使得识别过程也非常耗时。本文采用了鉴别力量分析方法提取Gabor图像特征中最有鉴别力的系数作为特征,这样极大地减少了Gabor图像特征的维数。另外,把经过Gabor变换得到的不同人脸图像中的同一尺度和方向的变换结果进行特征重组,得到多个新特征矩阵,每一新特征矩阵的贡献被本文所提出的自适应权重方法计算得到。最后对新特征矩阵抽取LDA特征进行识别。人脸识别实验显示了所建议方法能够有效地去除光照变化的影响。(3)研究了距离保持投影降维方法,提出一种改进的距离保持投影算法,并在此基础上进行了扩展。从某种意义上来说人脸是一种流形结构,人脸数据集是由某些内在变量控制所形成的非线性流形,如果在流形中能够找出光照、姿态、表情等控制变量,就可以极大地降低观测空间的维数。测地线距离是流形上两点之间距离最短的线,它能够描述人脸图像中由于光照、姿态、表情等变化而引起的非线性因素,而欧氏距离不能很好地度量由这些因素引起的人脸图像的非线性变化。因此,通过用测地线距离替代欧氏距离,本文提出了一种改进的距离保持投影算法。为了减少距离保持投影中邻域大小难以选取的问题,文中采用了一种对邻域大小不甚敏感的算法。针对距离保持投影流形学习算法没有充分利用样本的类别信息,不能用于分类,本文提出了一种基于距离保持投影的新的人脸识别算法。通过实验验证了本文方法可视化与分类识别能力。(4)研究了鉴别局部排列(DLA)特征提取算法,提出一种增强鉴别局部排列(EDLA)算法和核增强鉴别局部排列(KEDLA)算法。特征提取是人脸识别中关键的一步,所提取的特征必须对光照、表情、姿态等变换有较强的鲁棒性。DLA算法是一种基于局部最优和全局排列的特征提取算法,在人脸识别中获得了成功的应用。然而该算法识别性能严重依赖于参数的选择,对参数的选取极其敏感,并且该算法只利用了部分类别信息。为了提高算法的鲁棒性,本文提出一种对参数的选取不敏感且充分利用类别信息的增强DLA算法,并将此算法扩展到核空间,进而提出了KEDLA算法。实验结果表明,在光照变化条件下这两种算法的识别率要分别高于DLA、KDLA以及传统的子空间算法,说明了这两种方法对光照具有一定的鲁棒性。更多还原

【Abstract】 As motivated by the extensive potential applications in public security, financial security, human-computer interaction, etc, biometrics recognition has become one of the main topics in the fields of pattern recognition and artificial intelligence. Especially face recognition has become a most potential recognition technology by biometric characteristic for the merits of being natural, directly perceived, safe and convenient. During the past few decades, a lot of theories and algorithms have been studied and proposed, and face recognition under controlled conditions has been solved preliminarily. While face recognition under the uncontrolled, uncooperative conditions, is still a great challenge. There are a lot of non-controlled factors such as post variation, illumination variation, expression variation, and so on, among which the effects of illumination change is especially serious.Our work is focusing on the effect of illumination variation on the face image recognition. The emphases of the work are the face image preprocessing, feature extraction under varying illumination. The main contributions in this dissertation can be summarized as following.(1) A illumination subspace and quotient image method was studied, and then a novel method for varying illumination conditions was proposed.The 3D point source model in quotient method is too simple to effectively approximate arbitrary illumination conditions. Furthermore, this method might fail in obtaining the illumination invariant feature when there exists shadow in the input image. Therefore, an improved quotient image method based on 9D linear illumination subspace is presented to overcome these deficiencies. Firstly, the nine basis images of each subject, which can be used to synthesize face images of the object under arbitrary lighting conditions, are generated by means of one image of each object under the standard lighting conditions and the improved quotient images method. Then one new image of each subject under the same lighting conditions with an input image is synthesized by the corresponding basis images. Finally, based on the synthetic face images, face recognition under varying illumination can be performed.(2) Both illumination preprocessing method and facial feature extraction method based on Gabor wavelet transform were studied, and then an adaptive feature and weight selection method based on Gabor image for illumination-invariant face recognition was proposed.In order to alleviate the effect of illumination variations on face recognition, an illumination preprocessing algorithm based on local contrast enhancement is proposed. Gabor Wavelet exhibits the desirable characteristics of orientation selectivity and spatial locality, and it can capture images the edge features and local salient features at different directions and frequencies. Gabor Wavelet is often adopted to extract the robust features from face images. However, the high-dimensional Gabor features make recognition process computationally expensive and have high cost in space requirement. Therefore, an adaptive feature and weight selection method based on Gabor image for face recognition is proposed. Firstly,40 Gabor feature matrices which are reconstructed with the same scale and the same direction transform results of the different face images are obtained by regarding every Gabor wavelet transformed output image as an independent sample. Then the contribution of each new feature matrix can be adaptively computed by the proposed adaptive weight method. Thirdly, the Gabor feature coefficients which have more power to discriminate different classes than others are selected by discrimination power analysis to construct feature vectors. Lastly, linear discriminant analysis features are extracted to fulfill recognition task. Experiments on the face databases demonstrate its effectiveness.(3) Distance-preserving projection method based on manifold learning was studied, and then the improved distance-preserving projection method was proposed.Human face is a manifold structure in some sense. Face dataset is a non-linear manifold formed by some inner variables, such as illumination condition, facial expression, etc. if some controlled variables can be sought, space dimensionality could be reduced greatly. Geodesic distance is the shortest distance between the two points on manifold, and it can be described the nonlinear factors in face images caused by post variation, illumination change, expression variation, and so on. However, the nonlinear manifold of face images can not be well measured by the Euclidean distances when these factors change frequently. Replacing Euclidean distance by geodesic distance, the improved distance preserving projection algorithm is proposed. In order to choose a suitable neighborhood size effectively, the P-ISOMAP algorithm which is much less sensitive to the neighborhood size is used. Because distance-preserving projection algorithm can’t make full use of sample’s information and it is not suitable for classification. A novel face recognition method based on distance-preserving projection is proposed. And its effects have been verified in open face databases.(4) Face feature extraction method based on Discriminative locality alignment was studied, and then an enhanced discriminative locality alignment (EDLA) method and a kernel version of EDLA named Kernel EDLA (KEDLA) were proposed.Feature extraction is critical step in face recognition. The extracted feature should be robust to illumination, expression, poses, etc. Discriminative locality alignment (DLA) which is based on the idea of part optimization and whole alignment has better performance. However, the performance of DLA is too overly sensitive to the values of the parameters. In addition, it falls short of exploiting the full supervision information, since all points that belong to the same class can bring significant discriminant information to the dimensionality reduction matrix. Therefore, a novel supervised feature extraction method namely enhanced discriminative locality alignment (EDLA) is presented. EDLA is not sensitive on the choice of the parameters and both the local structure and class label information are fully considered in EDLA algorithm. Moreover, a kernel version of EDLA named Kernel EDLA (KEDLA) is developed through applying the kernel trick to EDLA to increase its performance on nonlinear feature extraction. The experimental results show that the two proposed methods have higher recognition rates than DLA、KDLA and the traditional subspace methods under different illumination conditions.更多还原