Embedded Manifold-Based Kernel Fisher Discriminant Analysis for Face Recognition

Manifold learning algorithms mainly focus on discovering the intrinsic low-dimensional manifold embedded in the high-dimensional Euclidean space. Among them, locally linear embedding (LLE) is one of the most promising dimensionality reduction methods. Though LLE holds local neighborhood information, it doesn’t fully take the label information and the global structure information into account for classification tasks. To enhance classification performance, this paper proposes a novel dimensionality reduction method for face recognition, termed embedded manifold-based kernel Fisher discriminant analysis, or EMKFDA for short. The goal of EMKFDA is to emphasize the local geometry structure of the data while utilizing the global discriminative structure obtained from linear discriminant analysis, which can maximize the between-class scatter and minimize the within-class scatter. In addition, by optimizing an objective function in a kernel feature space, nonlinear features can be extracted. Thus, EMKFDA, which combines manifold criterion and Fisher criterion, has better discrimination, and is more suitable for recognition tasks. Experiments on the ORL, Yale, and FERET face databases show the impressive performance of the proposed method. Results show that this proposed algorithm exceeds other popular approaches reported in the literature and achieves much higher recognition accuracy.