Based On Sparse Non-negative Matrix Decomposition And Radial Basis Function Neural Network Face Recognition Method

With the last decade, Face Recognition (FR) has become one of the most active and exciting research areas. One of the fundamental problems in face recognition is to find a suitable representation of the facial image. Most face recognition systems learn either holistic or parts-based representations. Principle Component Analysis (PCA), as a classical method for feature extraction, learns holistic representations of facial images, while Non-negative Matrix Factorization (NMF), a recently proposed approach, learns parts-based representations of faces. However, we argue that NMF can not only learn parts-based representations but also holistic ones with different sparseness constraints.Another issue in FR is to find an appropriate classifier. There are many conventional classifiers, which are generally divided into two categories: linear classifiers and nonlinear ones. Although successful in many cases, the linear classifier methods, such as k-nearest neighbor linear classifier, still fail to perform well when face images have larger variation in viewpoints, or with occlusions, which results in highly nonconvex and complex distribution. Radial Basis Function (RBF) neural networks, which can be used as non-linear classifiers, have been applied extensively in pattern classification because of their salient features in being universal approximators, possing the best approximation property and the fast learning speed and having more compact topology than other neural networks.In this thesis, we propose an efficient NMFs+RBF aggregate framework for FR, in which Non-negative Matrix Factorization with Sparseness Constraints (NMFs) is firstly applied to learn either the holistic representations or the parts-based ones by constraining the sparseness of the basis images, and then the RBF classifier is adopted for pattern classification.The comparative performances are studied among the NMFs+RBF method, the PCA+RBF method, and the PCA+FLD (Fisher's Linear Discriminant) method. All simulations are carried out on the ORL face database. The simulation results show that RBF classifier outperforms k-nearest neighbor linear classifier significantly in recognizing faces with occlusions, and the holistic representations are generally less sensitive to occlusions or noise than parts-based representations.