【作者】 郭金玉；

【导师】 苑玮琦；

【作者基本信息】 沈阳工业大学 ， 电工理论与新技术， 2009， 博士

【摘要】 生物特征识别是一项利用人类特有的生理或行为特征来进行身份识别的技术,它提供了一种高可靠性、高稳定性的身份鉴别途径。掌纹识别技术作为生物特征识别技术领域里的新成员,以其丰富的信息量、稳定而唯一的特征,近年来受到了全世界很多研究团队的重视,与此相关的新技术和新方法不断出现。如何从掌纹图像中有效地提取使之区别于其他个体的特征,是掌纹识别研究的一个关键所在。在众多特征提取方法中,子空间方法具有计算代价小、描述能力强、可分性好等特点,使用较少的特征向量就能取得较好的识别效果。本文利用PolyU掌纹图像库,以掌纹特征提取为研究目标,重点以子空间方法为研究手段,研究了一系列高性能的掌纹识别算法。具体来说,主要工作及贡献如下:(1)研究了实现掌纹方位无关性匹配的图像预处理算法。首先,对掌纹图像应用中值滤波器去除系统噪声,并应用阈值分割法分割背景区和目标区。然后,利用内部边界跟踪算法找到手掌边界,计算边界像素与手掌的底端任意点P_S之间的横坐标距离。根据横坐标距离曲线图中的两个局部最小值,自动确定两个指根部位置。最后,基于两个指根部建立参照坐标系,自动定位掌纹感兴趣区域(ROI)。(2)为了降低计算量,并且提取有利于分类的数据特征,研究了基于小波变换和部分最小二乘法(PLS)的掌纹识别算法。在建议的方法中,首先通过小波三级分解提取低频子图像,对低频子图像应用PLS提取掌纹子空间,然后将样本投影到掌纹子空间上,得到特征向量进行分类识别。(3)在掌纹等图像识别领域,Fisher线性判别(FLD)常会面临小样本问题导致的求解困难。在小样本情况下,为了采用FLD有效地提取分类特征,同时克服线性特征提取方法不能有效地提取像素间的非线性统计特性的缺点,研究了基于核主元分析(KPCA)和FLD相结合的新的掌纹识别方法。对每幅掌纹图像应用KPCA进行降维,然后将二维图像矩阵转换成一维图像向量。PolyU掌纹图像库中所有图像向量组成数据矩阵作为FLD的输入,提取分类特征,计算特征向量间的余弦距离进行掌纹匹配。(4)在小样本掌纹识别中,运用局部保持投影(LPP)方法时,特征方程矩阵存在奇异性。传统的解决方法是运用主元分析(PCA)获得原样本的低维特征子空间,在该空间中运用LPP进行特征提取。由于PCA和LPP的投影标准本质上是不同的,PCA降维时丢失许多重要的判别信息。另一方面,经典的PCA是基于一维向量进行降维的算法。将掌纹图像矩阵表示成向量的形式,破坏了图像像素间的原有空间结构关系,而这些空间结构关系对于分类来说是不可忽视的。为了解决这个问题,对传统的PCA+LPP方法进行改进,运用三种基于图像矩阵的方法——三级小波变换、图像下抽样和图像分块求均值降低掌纹图像维数,在低维图像上应用LPP提取局部结构特征,计算特征向量间的余弦距离进行掌纹匹配。(5)由于LPP是非监督学习算法,提取掌纹特征时没有考虑训练样本的类别信息,因而分类效果不够理想。为了获取判别意义上最优的掌纹特征,研究运用LPP和核直接判别分析(KDDA)相结合的新方法进行掌纹识别。在小样本图像识别中,为了解决特征方程矩阵的奇异性问题,首先运用图像下抽样降低掌纹空间的维数,然后应用LPP提取掌纹局部结构特征。LPP提取的特征作为KDDA的输入进行分类特征提取,计算特征向量间的余弦距离进行掌纹匹配。(6)非负矩阵分解(NMF)具有非负性和局部性的特点,是一种新型的特征提取方法。NMF是非监督学习算法,运用NMF提取掌纹特征时没有考虑训练样本的类别信息,因而分类效果不够理想。为了在提取掌纹非负局部特征的基础上能够很好地利用类别信息,本文采用非负矩阵分解和广义判别分析(GDA)相结合的新方法进行掌纹识别。为了降低计算量,在特征提取之前,应用小波变换对掌纹图像进行三级分解,提取低频子图像。在低频子图像上应用NMF+GDA提取掌纹非负局部分类特征,计算特征向量间的余弦距离进行掌纹匹配。更多还原

【Abstract】 Biometrics recognition is a kind of identification technology that uses the human’s special physiology or behavior characteristic, it provides a kind of high reliability, good stability approach of identity verificiation. As a new member of biometric family, palmprint has drawn great attentions from many research teams from all round the world since it has rich information, stable and uniqueness features. As a result, various palmprint recognition methods have been proposed.The key issue of a successful palmprint recognition approach is how to extract discriminant features from a palmprint image. Many feature extraction methods have been proposed. Among them the subspace methods have appealing properties, such as low time-consuming, good performance on expression and separation. It can obtain good recognition performance with less feature vectors. Palmprint images used for experiments come from PolyU database. This dissertation focuses on the feature extraction technologies based on subspace methods, and studies a series of efficient palmprint recognition algorithms. In detail, the main jobs and contributions are as follows:(1) This dissertation studies palmprint image preprocessing to match without the influence of orientation. First, traditional median filter is utilized to reduce system noises effectively, and threshold segment method is applied to palmprint images for image segment. Then, the inner border tracing algorithm is employed to find the palm border, and the horizontal distance between border pixels and any point P_s of the intersection line that is formed by the wrist and bottom margin of a palmprint image is calculated. There are two local minimums in distance distribution diagram. We can find two finger-web locations by the two local minimums. Finally, the coordinate system is founded based on the two fing-web locations. Palmprint region of interest (ROI) can be orientated automatically.(2) In order to reduce the computational quantity, and extract data feature which are propitious to classification, a palmprint recognition method based on wavelet transform and partial least square (PLS) is studied. A palmprint image is decomposed to low frequency sub-images by three-level wavelet transform. Then PLS is applied to get palmprint subspace. The original palmprint image was mapped into the subspace to get feature vector for classification.(3) In palmprint and other image recognition fields, Fisher linear discriminant (FLD) method often has no answer because of small sample size problem. In small sample size cases, in order to extract classification feature with FLD and overcome the disadvantage of linearity methods which cannot effectively extract nonlinear characteristics between bixels, a novel method for palmprint recognition based on kernel principal component analysis (KPCA) and FLD is explored here. In the algorithm, after the utilization of KPCA as a pre-processing step to reduce the dimensionality of a palmprint image, the 2D image matrix is then transformed into 1D image vector. FLD is used to extract classification feature vectors for all palmprint image vectors of PolyU palmprint database. Then the cosine distances between feature vectors are calculated to match palmprints.(4) In small sample size cases, when locality preserving projection (LPP) is applied to palmprint recognition, the matrix of the eigenvalue equation is singular. The traditional solution is to utilize the principal component analysis (PCA) as a pre-processing step aiming to reduce the dimensionality of the palmprint space, then LPP is applied to extract feature. Since the projection criterion of the PCA and that of LPP are essentially different, the pre-processing step applied the PCA to reduce the dimensionality could result in the loss of some important discriminatory information. On the other hand, classical PCA reduces the dimension based on 1D vector. Palmprint image matrixes are expressed as vector, which destroys the original space structure relations between bixels. These space structure relations are not neglectable for classification. To solve the above problems, and improve convetional PCA+LPP, three methods based on image matrix, the three-level wavelet transform, image down-sample, and the mean of block segmentation, are presented to reduce palmprint space dimensionality. Then LPP is used to extract the local structure features. The cosine distances between feature vectors are calculated to match palmprints.(5) LPP is an unsupervised learning method, and it doesn’t consider class information of samples applied to extract palmprint features, so the classification performance is not ideal. In order to obtain the best palmprint features in discriminant point of view, a new palmprint recognition method based on LPP and kernel direct discriminant analysis (KDDA) is explored. In small sample size image recognition, in order to resolve the singularity of the eigenvalue equation, image down-sample is first applied to reduce the palmprint dimensionality, and LPP is then used to extract the local structure features. The local features are the input of KDDA to extract classification features. Then the cosine distances between feature vectors are calculated to match palmprints.(6) Non-negative matrix factorization (NMF) has non-negative and local characteristics, and it is a new feature extraction method. NMF is unsupervised learning method, and doesn’t consider class information of samples applied to extract palmprint features, so the classification effect is not ideal. In order to use class information better after the negative and local features of images are extracted, a new palmprint recognition method based on non-negative matrix factorization and general discriminant analysis (GDA) is proposed. Before extracting features, the three-level wavelet transform is utilized to palmprint images to get the low resolution sub-images. Then NMF and GDA are applied to extract non-negative and local palmprint classification features. Then the cosine distances between feature vectors are calculated to match palmprints.更多还原