Local Fisher Discriminant Analysis with Maximum Margin Criterion for Image Recognition

Reducing the dimensionality of data without losing intrinsic information is an important preprocessing step in image recognition. Local Fisher Discriminant Analysis (LFDA) is a linear projective map that arises by solving the multimodal problem, which effectively combines the ideas of FDA and LPP. However, since the limited data pairs are employed to determine the discriminative ability, such local discriminative methods usually suffer from the maladjusted learning. To improve the discriminant ability of LFDA, this paper proposed an improved manifold learning method, called local and global marginal discriminant analysis (LGMDA), by incorporating the maximum margin criterion (MMC) for image recognition. As a result, the proposed method tries to find the sub manifold that best discriminates different classes and preserves the intrinsic relations of the local neighborhood in the same class according to prior class information. Experiments on the COIL-20 and YaleB images databases show the effectiveness of the proposed LGMDA.