| In past decades, the research on face recognition has become one of the most hot research topics on pattern recognition and computer vision, and the technology of face recognition is not only the needs of the development of science but also a necessary condition for building a harmonious society. Based on the assumption that face images from the same subject lie in a linear subspace, to use a linear representation model has become the mainstream for solving issues of face recognition. To classify a probe, we first solve the model to obtain the representation parameters (i.e. coding vector) and then calculate the distance between probe and each subspace. The decision is made in favor of the minimal distance. However, when big changes or large noises appear in facial image, the location of it in feature space will be hard to be predicted, which influences the recognition result seriously. In order to solve this problem, a competitive representation classification method is proposed. It seeks for the subspace where the probe lives, which improves the abilities of representation and recognition for the probe. Furthermore, many previous face recognition methods treat the features of facial image independent and ignore the correlation among them. So this thesis focuses on this research issue and proposes a matrix-based representation classification method for face recognition. Overall, both two methods improve the performance of face recognition in noise cases.The highlights of the two proposed methods are described as follows:1. A competitive sparse representation classification method (CSRC) is proposed. In the process of multi-step iteration, the lowest competition-elimination mechanism enhances the competitiveness for representing the probe. At the same time, the sparse representation parameters are obtained which is benefit for classification.2. A fast solution is adopted to optimize CSRC model, which improves the recognition speed greatly.3. A matrix-based representation classification (MRC) for face recognition is proposed. Since the correlation between face features are taken into the consideration which is helpful to reduce the influence by noise, the proposed model has strong stability to recognize a probe correctly even the probe is occluded seriously.4. A self-adaptive classifier is adopted to avoid the influence by noise in classification stage. In the process of recognition, MRC dynamically determines the threshold differentiates representation residuals and noise error when training images and testing image changed. Then, the classification result will be more stable and reasonable. |
