Research On Subspace Analysis Based Face Recognition

Face recognition is one of the important biometric identification technologies. The main research topic on face recognition is how to make computer identify a specific person. The key issue of a successful face recognition approach is how to extract discriminant features from a face image. Many feature extraction methods have been proposed. Among them the subspace methods have been the most popular approach owing to their appealing properties, such as low time-consuming, good performance on expression and separation. This dissertation focuses on the feature extraction technologies based on subspace methods. The contributions of the dissertation are:1) The features extracted with IMPCA algorithm have very high dimensionality. To overcome this shortcoming, two-direction IMPCA (2DIMPCA) was proposed. 2DIMPCA performs IMPCA twice: one in horizontal direction and the other in vertical direction. 2DIMPCA not only has good performance, but also reduces the feature dimensionality.2) Combing the image matrix model and LPP, a new face recognition method called Image Matrix LPP (IMLPP) was proposed. Like general face recognition algorithms, LPP is based on vectors. Their first task is to convert image matrix into vector. IMLPP works directly with images in their native state, i.e. two dimensional matrices. IMLPP keeps the raw special position information of pixel in face image and has good recognition rate. However, it is a time-consuming method for its iterative processing.3) OLPP is a face recognition algorithm based on manifold learning, and it can extract nonlinear orthogonal features. However, OLPP is based on iterative processing so the algorithm is complicated. OLPP belongs to unsupervised methods and does not make full use of the labels' information of samples. In this dissertation, a new face recognition method based on orthogonal discriminant locality preserving projections (ODLPP) was proposed. Based on LPP, ODLPP takes into account the between-class information, changes the objective function, and then orthogonalizes the basis vectors of the face subspace. Experimental results indicated the promising performance of the proposed method.4) Combing the idea of nonlinear kernel mapping and LPP, a new face image feature extraction and recognition method based on kernel supervised locality preserving projections (KSLPP) was proposed in this dissertation. KSLPP projects the samples into high-dimensional feature spaces by some nonlinear mapping, combines the face manifold local structure information and the labels' information, and extracts the nonlinear features of a face for recognition. The experimental results showed that KSLPP had good performance for face feature extraction and recognition.5) One training sample problem is the case which we can not ignore for real application. In this dissertation, we proposed an approach to make DCV method applicable when a person has only one training image. Our approach is based on the assumption that human faces share similar intrapersonal variations. The intrapersonal variations of a training set can be estimated from the collected generic face set.