Research On Subspace Analysis-based Palm-dorsa Vein Recognition Methods

As an important part of the biometrics, palm-dorsa vein recognition has a very wide application prospect. How to extract discriminant features from a palm-dorsa vein image is a key issue of a successful palm-dorsa vein recognition approach. Many feature extraction methods have been proposed. Among them the subspace method is the most popular approach owing to its property, such as good expression, low time-consuming and good performance on separation. In this dissertation, the use of subspace analysis for feature extraction from the palm-dorsa vein images is investigated. In order to resolve existing feature extraction problems, we combine several feature extraction methods and propose a series of efficient palm-dorsa vein feature extraction algorithms.The main creative work and contributions of the dissertation are:(1) The dissertation extends the applicability of the partial least square which is usually used for linear regression to palm-dorsa vein recognition for the first time. Since partial least square extracts the principal components from the input variable (the input variable is the training sample set) and the output variable (the output variable is the classification variable of training samples) simultaneously by utilizing the correlation between the extracted principal components, it can extract the features which contain more classification information of the images and are very propitious to matching. The experimental resuls show that compared with the traditional dimensionality reduction methods-principal component analysis and independent component analysis, partial least square has better performance for palm-dorsa vein feature extraction and recognition.(2) Independent component analysis doesn’t make use of classification information of training samples when the palm-dorsa vein features are extracted, so the classification performance is not ideal. Combing general discriminant analysis and independent component analysis, a new palm-dorsa vein recognition method based on general discriminant independent component analysis is proposed in this dissertation. Independent component analysis is firstly used to reduce palm-dorsa vein space dimensionality and the statistically independent basis vectors of training samples are gained. In the subspace spanned by the statistically independent basis vectors, general discriminant analysis is applied to extract the discriminant features.The experimental results show that the integration of discriminant information and global features greatly improves the recognition rate of the algorithm.(3) As an effective feature extraction method, non-negative matrix factorization decomposes the non-negative sample data under the non-negative constraint. The basis vectors learned by non-negative matrix factorization contain the local information and have the clear physical meaning. However non-negative matrix factorization is an unsupervised learning method and recognition performance is not good. Combing kernel direct discriminant analysis and non-negative matrix factorization, a new palm-dorsa vein recognition method based on kernel discriminant non-negative matrix factorization is proposed to resolve this case. The proposed method constructs the projection vectors using the local features of the training samples, so that the projection vectors of samples come from same class are centralized and the projection vectors of samples come from different classes are dispersed in the new feature space.The experimental results suggest that the proposed kernel discriminant non-negative matrix factorization algorithm provides good representation and achieves ideal recognition accuracy in palm-dorsa vein recognition.(4) The dissertation extends the applicability of feature extraction methods based on manifold learning to palm-dorsa vein recognition. Locality preserving projection is applied to preserve the manifold structure and extract the local discriminant features. But in small sample size cases, when locality preserving projection is applied to recognition, the matrix of the eigenvalue equation is singular. The traditional solution is to utilize the principal component analysis to reduce the dimensionality of the original data samples and ensure the matrix of the eigenvalue equation is nonsingular, then locality preserving projection is applied to extract feature. However, principal component analysis only considers the second-order statistics of the date and could not utilize the important information contained in the high-order relationships between pixels. To solve the above problem, kernel principal component analysis instead of principal component analysis is presented to reduce palm-dorsa vein space dimensionality so that the nonlinear features are extracted. The experimental results show that the algorithm combined kernel principal component analysis and locality preserving projection is powerful than the algorithm combined principal component analysis and locality preserving projection.(5) The palm-dorsa vein recognition methods based on two-dimensionl feature are studied in this dissertation. The traditional methods based on two-dimensional image matrix extract features only in the horizontal direction of the image and feature’s dimensionality is high. A new palm-dorsa vein recognition method based on block segmentation and double two-dimensional locality preserving projection is proposed to resolve this case. In the proposed method, block segmentation is firstly utilized as a pre-processing step. Then two-dimensional locality preserving projection is executed to the set composed of all the subimages which are respectively in the^ne location of all the training images for extracting the local features in horizontal direction.Transpose the feature matrix and execute two-dimensional locality preserving projection again, the local features in vertical direction are extracted. After projecting the subimage in each position into the corresponding subspace respectively, we will gain the feature matrix of the image.The algorithm executes two-dimensional locality preserving projection in both horizontal and vertical directions, so that the feature’s dimensionality in two directions is effectively reduced. The experimental results suggest that compared with the traditional feature extraction methods based on two-dimensionl matrix, block segmentation and double two-dimensional locality preserving projection algorithm retains more spatial structure of the image pixels, reduces more influence caused by rotation and illumination, and achieves higher accuracy in palm-dorsa vein recognition.