节点文献
四种人脸识别方法研究
Study on Four Face Recognition Methods
【作者】 何光辉;
【导师】 唐远炎;
【作者基本信息】 重庆大学 , 计算机应用技术, 2010, 博士
【摘要】 人脸识别是近二十年来模式识别中的一个重要课题,近年来人脸识别系统已经可以较为准确的在某些限定的条件下对人脸进行识别,随着应用范围的不断扩展,受到越来越多企业和单位的重视。现在,人脸识别已经在公共安全、智能监控、数字身份认证、电子商务、多媒体和数字娱乐等领域显现了巨大的应用价值。随着人脸识别的应用越来越广泛,人们对于人脸识别实用性的要求则变得越来越严格。虽然现今的大多数人脸识别算法已经可以在较为固定的环境下对人脸进行较为正确的识别,但实际使用过程中,由于受到光照、姿势、表情、遮挡等因素的影响,人脸识别的精度仍然不能满足实际的要求。本文针对上述存在问题进行探究,并取得了如下的主要成果:1、针对主成分分析( PCA)方法和仿生方法在特征提取和降维方面的不足,提出一种Gabor特征提取的仿生人脸识别方法。该方法首先提取人脸图像Gabor特征向量,经2DPCA方法降维处理后运用仿生识别方法对其进行人脸识别。在Yale Face database B和PIE人脸库上验证了该方法的有效性。实验结果表明,该方法的分类准确性优于仿生识别方法和PCA等方法。2、针对基于欧氏空间的人脸识别算法框架与人类的视觉感知系统有着很大差异的问题,提出了一种相关性子空间人脸识别方法,通过相关性度量多维尺度分析(Correlation MDS,CMDS)方法寻找到一个相关性保持的子空间,将高维人脸数据投影到此子空间中,使得人脸图像间的相似性能够得到保持。既然高维数据中的非线性几何结构常常是嵌在数据间的相似性,因此相关性子空间人脸识别方法能够有效地获取高维人脸数据中的非线性流形结构。在多个数据集上的实验结果表明:该方法可有效地提取高维数据的内在结构。3、从人类认知方式出发,提出了一种基于统计学习的局部匹配人脸识别方法,该方法首先将人脸图像划分成若干小块,将每个子块看成一弱分类器,接着利用Adaboost学习算法将这些弱分类器组成一个强分类器,将各个子块(特征)有效地组合起来,发挥他们的最佳判别能力,提高最终的分类效果。与整体匹配方法相比,局部匹配的人脸识别对人脸局部变化(光照、表情、姿态等)更具有健壮性,实验结果表明该方法可有效地提高人脸的识别准确率并对人脸的表情和光照具有较好的鲁棒性。4、针对稀疏表示方法存在对负系数缺乏必要的物理意义解释且不能通过常用的梯度下降法来求解的问题,提出了一种非负稀疏表示的人脸识别方法,其理论基础是将测试图像表示成训练图像的非负稀疏线性组合,这样更符合人类的认知且更具有实际的物理意义,还可将稀疏表示中L1范数最小化问题转换成基于L2范数最小化的稀疏表示问题,从而能够通过常用的梯度下降法来求解。该方法无需进行降维,特征选择,合成训练样本和变换域等操作。在Yale B扩展数据库上的大量的实验结果表明了该算法的有效性。
【Abstract】 Face recognition was one of the most prominent areas in pattern recognition for several decades. Numerous face recognition methods have been proposed. Current systems can do fairly accurate recognition under constrained scenarios using these face recognition methods. It has attracted much attention due to its potential application values. Nowadays, face recognition has great value in public safety, intelligence surveillance, identification, E-commerce, multimedia, digital entertainment and so on.Requirements of the practicality of the face recognition are increased along with the wide use. Also, most of the current face recognition methods can do fairly accurate recognition in fixed environment. But the fixed environment constrain is not suitable for most of the application. This paper discuss the problems mentioned above, the following contributions are made:1. In order to overcome the shortcomings of Principal Component Analysis(PCA) and bionic methods in feature extracting and dimension reduction, a method for extracting Gabor features of face images based on Gabor wavelet is presented. First, Gabor features are extracted from face images. After reduced dimension by 2DPCA, the nearest classifier is trained for classification. The experiments being performed on Yale Face Database B and PIE human face image database show the method presented in this paper is superior to bionic recognition algorithm and PCA algorithm.2. According to the difference between face recognition algorithm in Euclidean metric subspace and human visual perception system, we propose a face recognition method in correlation measure subspace. It aims to find a subspace with preserving correlation measure by correlation multidimensional scaling. Then high-dimensional face image data are projected into such a space. As a result, the similarities between face images are preserved. Since there is a low-dimensional nonlinear manifold lying in the high-dimensional data, the dimensionality and nonlinear geometry of that manifold often is embedded in the similarities between the data points. Thus, the proposed method can effectively learn the nonlinear manifold embedded in the high-dimensional face images. The experimental results on various databases show that the proposed methods can effectively extract the intrinsic structure of high-dimensional data.3. From the way of human cognition, we propose a local matching face recognition method based statistical learning. It first divides the image into several subimages, then each subimage is considered as a weak classifier. The Adaboost learning algorithm is used to train the weak classifiers and construct a strong classifier. As a result, each subimage is effectively combined together to explore their best discriminanting power and improve the classification accuracy. Compared with holistic matching methods, the local matching method is robust to variations in illumination, expression, and pose, etc. The experimental results show the proposed method can improve the face recognition accuracy and is robust to variations in illumination and expression.4 Sparse representation is difficult to explain the physical meanings of negative coefficient and can not be solved by using gradient descent method. According the limitation of sparse representation, we propose a face recognition method via nonnegative sparse representation. In this method, the testing image is represented as a linear combination with nonnegative coefficients of the training images. The nonnegative limitation is more consistent with human cognitive and has practical physical interpretation. The proposed method transforms L1 norm minimization into L2 norm minimization which can be solved by using gradient descent method. This method achieves state-of-the-art performance using raw imagery data, with no need for dimension reduction, feature selection, synthetic training examples or domain-specific information. Extensive experiments on Extended Yale B database verify the efficacy of the proposed method.
【Key words】 Bionic face recognition; Correlation measure; Image partition; Sparse representation; Random corruption;