Research On Subspace Feature Extraction Methods For Face And Palmprint Recogniton

This thesis is focused on some issues related to subspace feature extraction methods for face and palmprint recognition. These issues mainly include novel subspace analysis methods derived from the idea of manifold learning, the theoretical framework of graph embedding for matrix-based feature extraction algorithms with their applications to face and palmprint recognition. In addition, in this thesis the fusion of the subspace features of face and palmprint at the feature extraction level is also discussed. The main contributions can be exhibited by the following aspects:1. As an alternative subspace method, Locality Preserving Projections (LPP) algorithm has been proposed recently, which takes into account the space structure of the samples, and in the process of dimension reduction, it can thus find a good linear embedding that preserves local structural information and intrisinc geometry of the data space. However, in practice for the LPP algorithm, a two-step framework (Principal Component Analysis + Locality Preserving Projections) is required, which is indirect and uncomplete. To attack these problems, in this thesis a Direct Locality Preserving Projections (DLPP) algorithm is proposed. This algorithm solves locality preserving problem via simultaneous diagonalization, and can avoid the singularity of the matrices. This algorithm accepts high-dimensional raw images as input, and optimizes locality preserving criterion directly, without any dimensionality reduction steps. Experimental results on the PolyU palmprint database and the ORL face database show the effectiveness of the proposed algorithm.2. The LPP algorithm is based on vector-space model. Under this model, the original two-dimensional image data are reshaped into a one-dimensional long vector by rows or columns, which leads to not only the loss of some structural information residing in original 2D images but also a too high-dimensional data space. Inspired by the idea of two-dimensional Principal Component Analysis (2DPCA) algorithm, in this thesis a novel algorithm, two-dimensional Locality Preserving Projections (2DLPP) algorithm, is proposed, which is a straightforward manner based on locality preserving criterion and the image matrix projection. This algorithm directly projects the image matrix under a specific projection criterion, rather than using the stretched image vector. Experimental results on the Yale face database and the PolyU palmprint database show that the 2DLPP algorithm outperforms the conventional Principal Component Analysis (PCA), vector-based LPP and two-dimensional Principal Component Analysis (2DPCA) algorithms in terms of the recognition accuracy rates.3. As the LPP model is linear, it may fail to extract the nonlinear features. To attack this problem, the Kernel Locality Preserving Projections (KLPP) algorithm proposed in this thesis is to nonlinearly map the data into a feature space in which the dataset has a linear structure or a structure as linearly separable as possible, then LPP is performed in feature space to find the locality preserving projection vectors for final classification. Kernel tricks are used to change the problem of implementing KLPP algorithm in feature space into a problem of performing LPP in the Kernel Principal Component Analysis (KPCA) transformed space. Experiments on the ORL face database and the PolyU palmprint database demonstrate the effectiveness of the proposed algorithm.4. Recently proposed matrix-based methods in the research community have been shown to be effective ways to avoid the problems of high dimensionality and small sample sizes that are associated with vector-based methods. In this thesis, a general framework for matrix-based feature extraction algorithms is proposed from the point of view of graph embedding. It is found that through designing meaningful graph structures which satisfy various objective functions, this framework can be used as a platform to derive new matrix-based algorithms, in this direction, a novel matrix-based algorithm, i.e. two-dimensional Discriminant Embedding Analysis (2DDEA). is proposed, which explicitly takes into account the matrix-based intra-class submanifold and inter-class submanifold by integrating the local intra-class compactness information and the non-local inter-class separability information. The proposed algorithm does not require any hypotheses on the distribution of the dataset and is thus a simple data-driven approach. It is also shown that 2DLDA is actually a special case of the proposed 2DDEA method. Experimental results on three public image databases show the effectiveness of the proposed algorithm.5. In practical applications, due to the comlexity and the unpredictability of the working environments, biometric systems based on single trait have exhibited several unresolved problems. Therefore, in this thesis the subspace features of face and palmprint are fused at the feature extraction level for personal identification. Two well developed subspace methods, i.e. Principal Component Analysis (PCA) and Independent Component Analysis (ICA) are used to extract the face and palmprint features. It is found that the performance is significantly improved in both cases, especially in the case of feature fusion using ICA, encouraging results with a 99.17% recognition accuracy rate using a test set sized of 40 people are obtained. The results of this work suggest that a multimodal system integrating of faces and palmprints can offer substantial performance gain that may not be possible with a single biometric indicator alone.