【作者】 李勇智；

【导师】 杨静宇；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 特征抽取是模式识别中最基本的问题之一。无论是人脸识别还是字符识别,提取有效的鉴别特征是解决问题的关键。本文就有关几种线性和非线性特征提取方法的理论与算法进行了研究,并且所提出的一些特征提取的新算法在人脸识别方面得到了较成功的应用。首先,本文针对非矩阵分解的理论,提出了具有正交性的投影轴的计算方法和具有统计不相关性的投影轴的计算方法。这种处理方法的目的为了减少低维空间中投影轴之间的统计相关性,提高识别率。实验结果表明提出的两种特征提取方法在识别率方面整体上好于原非负矩阵分解特征提取(NMF)方法。针对非负矩阵分解方法在特征提取过程中没有充分利用训练样本的类别标签信息,本文提出一种新的有监督的非负矩阵分解方法,这种方法的特点:一是它充分直接利用训练样本的类别信息,二是在计算上仍然采用与非负矩阵分解方法相同数学公式,这种新特征提取方法被称为组合类别信息的非负矩阵分解(CINMF)方法。其次,针对非线性特征提取问题,基于核技术的理论,本文给出一种监督化的KPCA方法,即组合类别信息的核主成分分析(CIKPCA)。由于核主成分分析(KPCA)是无监督学习方法,在特征提取过程,KPCA中不能充分利用训练样本的类别信息,而CIKPCA则是克服这一弱点。在分类时本文采用基于两种特征融合的分类策略进一步改进CIKPCA方法的识别率。实验结果表明提出的新方法在识别率方面整体上超过常用的核主成分分析(KPCA)方法,在某些人脸数据库上,CIKPCA甚至超过了KLDA。另外,基于(核)最大间距准则,本文提出了一组具有统计不相关性的最优(核)鉴别矢量集的计算方法。新的方法的目的是消除了特征空间上最优(核)鉴别矢量间的统计相关性,提高了特征提取的有效性。最后,基于流形学习的理论,本文提出一种新的无监督的鉴别投影方法,这种新的方法是基于样本的局部和非局部统计量而建立的映射,它的鉴别准则是通过特征矢量的非局部散度与局部散度之间差的最大化来刻画的,准则目的是使得投影后特征矢量的非局部散度最大化,同时也使局部散度最小化。这种方法被称为最大间距鉴别投影(Marginal Discriminant Projection,MDP)。通过在ORL人脸库和AR人脸库上进行实验,比较了MDA、LDA、局部保持投影(LPP)和无监督鉴别投影(UDP)四种方法的识别率。另外,基于流形学习理论,本文提出一种新的针对图像矩阵的维数压缩方法。这种方法是基于样本图像矩阵来构建非局部散度矩阵和局部散度矩阵的,并且通过引入邻接矩阵来刻画高维数据的局部几何结构。准则函数是投影样本的非局部散度与局部散度之商的最大化来刻画的。新的方法是对图像矩阵行和列方向同时进行维数压缩而得到特征矩阵。这种方法被称为基于图像矩阵的双方向无监督鉴别投影((2D)~2UDP)。通过在ORL人脸库和AR人脸库上进行实验,结果表明本文提出的新方法在识别率方面整体上好于基于图像矢量PCA、LPP、UDP和(2D)~2PCA。更多还原

【Abstract】 Feature extraction is the elementary problem in the area of pattern recognition. It is the key to solve the problems such as face identification and handwritten character recognition. In this paper, we focus our attention on linear and nonlinear feature extractions and develop some new algorithms as regards it. And, these algorithms are verified to be effective in the application of image identification.Firstly, based on the Non-negative matrix factorization (NMF), a new algorithm of orthogonal projection axis and a new algorithm of statistically uncorrelated projection axis for feature extraction are proposed in this paper. Aim of the proposed methods is reducing or eliminating the statistical correlation between features and improving recognition rate. The experimental results on Olivetti Research Laboratory (ORL) face database and YALE face database show that the new methods are better than original NMF in terms of recognition rate.Non-negative matrix factorization (NMF) is an unsupervised feature extraction method in image recognition, meaning that NMF does not sufficiently use the class information of given training sample in feature extraction. A novel supervised feature extraction method based on non-negative matrix factorization is presented in this paper. The new method has two traits: one is to sufficiently utilize a given class label of training sample in feature extraction and the other is to still follow the same mathematical formulation as NMF, so the new feature extraction method is named class-information-incorporated non-negative matrix factorization (CINMF). Besides, in order to further improve recognition rate, the paper presents a new classification strategy based on fusion of two kinds of feature vector.Secondly, for nonlinear feature extraction, a novel supervised feature extraction method based on kernel principal component analysis (KPCA) is presented in this paper. The method is named as class-information-incorporated kernel principal component analysis (CIKPCA). As a nonlinear feature extraction, the conventional KPCA is an unsupervised method, it is not to sufficiently utilize a given class label information of training kernel sample in feature extraction, but CIKPCA overcomes the drawback. The paper presents a new classification strategy by fusion of two kinds of feature vector in order to further improve recognition rate. The experimental results show that the new method is better than KPCA in terms of recognition rate, and even outperforms KLDA. Besides, based on the (kernel) maximum margin criterion, new algorithms of statistically uncorrelated optimal (kernel) discriminant vectors for feature extraction is presented in this paper. The proposed methods have more powerful capability to eliminate the statistical correlation between features and improve efficiency of feature extraction.Lastly, based on manifold learning, a new unsupervised discriminant projection for dimensionality reduction of high dimensional data is presented in this paper. The new projection can be seen as a linear approximation of a multimanifolds-based learning framework which is based on both the local and nonlocal statistically quantities. The discriminant criterion function be characterized by difference between the nonlocal scatter and the local scatter, seeking to find a group of projection axis that simultaneously maximizes the nonlocal scatter and minimizes the local scatter of feature vector. Locality preserving projection (LPP) considers only the local scatter for classification . The experimental results on Olivetti Research Laboratory (ORL) face database and AR face database show that the proposed method consistently outperforms LPP and UDP, and even outperforms Fisher linear discriminant analysis (LDA).To avoid the complication of a singular local scatter matrix, we present a new feature extraction method by the idea of manifolds learning, the trait of the method is to exploit image matrixes to directly construct local scatter matrix and nonlocal scatter matrix. Its criterion function is characterized by maximizing the ratio of the nonlocal scatter to the local scatter after the samples are projected. The advantage of this new approach is that the reduction in dimensionality can be achieved in both row and column directions; the method is called the two-directional two-dimensional unsupervised discriminant projection (i.e. (2D)~2UDP). The experimental results on ORL databases and AR databases indicate that the new method is the highest among LPP, PCA, (2D)~2PCA and (2D)~2UDP in terms of recognition rate.更多还原