Kernel semi-supervised marginal fisher analysis and its application to face recognition

In the recent years, the manifold learning methods, which attempt to project the original data into a lower dimensional feature space by preserving the local neighborhood structure, have received much attention within the research communities of image analysis, computer vision and document data analysis. Among them, the recently proposed marginal fisher analysis (MFA) achieved high performance for face recognition. However, MFA is still a linear technique and usually deteriorates when labeled information is insufficient. In order to resolve those problems, we propose a kernel semi-supervised marginal fisher analysis (KSMFA) which not only exploits the nonlinear features but also preserves the global structure of labeled and unlabeled samples in addition to separating labeled samples in different classes from each other. Experimental results on the face databases indicate that the proposed KSMFA method is more effective than the MFA method and some existing kernel feature extraction algorithms.