Research On Face Recognition Methods With Subspace Linear Projections

Since face recognition is widely applied in many domains, the researches on it have been the highlight of the research in pattern recognition field at home and abroad. Subspace linear projection method is one of the mainstream techniques for face recognition. It has been given more and more attention owing to its good properties such as strong describing ability, efficient computation and easy implementation. We have devoted to theoretics and algorithms about the subspace linear projection methods and have made several conclusions which help us deeply comprehend the problem of face recognition and have developed a new algorithm,the algorithm is verified to be effective in the application of face recognition. The works that I have done are as follows:Firstly, in this paper bilinear interpolation method is used to scale images, which was simple and effective. In addition, the mean variance normalization method can eliminate the negative effects of illumination. The combibation of these two preprocessing steps is implemented in PCA-based face recognition, and the better results achieved in our experiments.Secondly, as well known, Principal Component Analysis(PCA) and Linear Discriminant Analysis(LDA) are linear projection methods. PCA seeks directions that are efficient for representation, LDA seeks directions that are good at discriminating samples. In the paper, the two face recognition techniques are discussed and compared. The experiments are done in the ORL face databases, and the two methods above are employed to extract features. The results tell that LDA obtains better recognition performance comparing with PCA.Thirdly, the basic principle of Independent Component Analysis (ICA) and Support Vector Machine (SVM) are studied, including their advatages and disadvantages respectively. On the basis of SVM muliti-classification, an improved binary tree support vector machines algorithm is proposed. A new face recognition approach based on Independent Component Analysis and Binary Tree Support Vector Machine is used in this paper. Here ICA is used to extract face image feature, then the recognition using binary tree support vector machine is carried out. Also, the simulation experiments are implemented in the ORL database.Experiment results demonstrate that this method can obtain good recognition performance. Finally, as a method which can extract parts-based feature, Non-negative Matrix Factorization (NMF) has been already applied in face recognition successfully. In this paper, the theory and characteristic with NMF are explored, and the two improved algorithms——Local Non-negative Matrix Factorization (LNMF) algorithm and NMF with Sparseness Constraints (NMFs) algorithm are analyzed and imported for face recognition. It is known that LNMF algorithm and NMFs algorithm are feasible according our experimental results and recognition accuracies are improved to some extent.