Research On Subspace Analysis-based Feature Extraction And Face Recognition

Face recognition continues to be a hot topic in pattern recognition field due to its wide range of applications such as commercial and law enforcement applications. A central issue to a successful approach for face recognition is how to extract discriminant feature from the facial images. Many feature extraction methods have been proposed and among them the subspace analysis has received extensive attention owing to its appealing properties. Now the subspace analysis method has been the most popular technology for feature extraction and face recognition. The dissertation investigated the use of subspace analysis for feature extraction from the facial images and recognition. The main contributions of the dissertation can be noted as following: 1. The dissertation gave a detailed analysis on the singular values (SVs) of facial images using projection analysis and then the reason why the SVs are not enough for face recognition was revealed. Based on this observation, a new algebraic feature was proposed by using singular value decomposition and projective method. Its robustness was also proven. The experimental results on the standard ORL face databse and Yale face database demonstrate that in comparision with the traditional SVs, the proposed algebraic feature contains more useful information in a smaller dimensionality and is a robust and more effective algebraic feature. 2. The dissertation extended the applicability of the latest heteroscedastic LDA which is based on the Chernoff criterion (HCLDA) to face recognition for the first time. As the HCLDA is able to simultaneously extract the discriminant information present in the differences between per class means and the differences between per class covariance matrices, it should be superior to the tradition Fisher criterion-based LDA in theory. However, in HCLDA the total within-class scatter matrix and per class within-class scatter matrices are required to be full rank, which is seldom satisfied in the face recognition tasks. In order to overcome this problem, the dissertation first improved the traditional PCA by introducing a named maximum generalized Fisher value principal component selection (MGPCS) strategy and then used the improved PCA for dimensionality reduction to make the total within-class scatter matix nonsingular. Furthermore, the latest maximumum entropy covariace selection strategy was selected to estimate the per class within-class scatter matrices. The experimental results show that during the dimensionality reduction stage, the improved PCA based on the prosoped MGPCS strategy can reserve more discriminant information than the traditional PCA and thereby improving the final classification results. 3. The dissertion proposed a novel variant on LDA that was referred to as relevance weighted uncorrelated LDA or RWULDA by integrating the relevance weighted Fisher critierion and uncorrelated LDA. The RWULDA method can not only restrain the negative influenence of the so-called outliner classes on the derivation of the optimal discriminant vectors, but also guarantee the obtained discriminant feature components are statistically uncorrelated. The experimental results on two subsets from the AR face dataase and FERET face database demonstrate the promising performance and effectiveness of the proposed technique. 4. After studied three typical variants on LDA, i.e. EFM, DLDA and NLDA, which are proposed to address the well-known small sample size problem, the dissertation revealed that the discriminant features derived from EFM, DLDA and NLDA are all statistically uncorrelated. In addition, the orthogonality of the discriminant vectors derived from NLDA was also revealed. Furthermore, the dissertation performed a detailed analysis on the selected subspace of each LDA methods and then gave the answer to the question why NLDA always outperforms than the other two methods. In order to simultaneously solve the small sample size problems and weaken the dominant influence of outlinear classes, the dissertation extended the NLDA technology by integrating the weighting scheme and uncorrelated LDA. The resulting method was referred to as weighting uncorrelated NLDA or WUNLDA. The efficiency and superiority of WUNLDA were demonstrated by the experiments on two subsets from the AR face database and FERET face database respectively. 5. In the dissertation, motivated by the success that SVM, kernel PCA and kernel FDA have in pattern classification tasks, the proposed WUNLDA was generalized to nonlinear WUNKDA method by integrating kernel method. Obviously, the novel WUNKDA method retains all merits of the WUNLDA method, while being able to extract the nonlinear feature. The new WUNKDA algorithm was tested, in terms of the simplified ability and recognition accuracy, on a more complicated subset from the FERET face database. The experimental results indicate that the proposed methodology is not only able to simplify the distribution of the face patterns, but improves the final classification results.