A novel Gaussian probabilistic generalized 2DLDA for feature extraction and face recognition

In this paper, the generalized two-dimensional Fisher´s linear discriminant (G-2DFLD) method is extended by incorporating Gaussian probability distribution information into the definition of the between-class and within-class scatter matrices to develop a novel Gaussian probabilistic generalized two-dimensional linear discriminant analysis (GPG-2DLDA). A Gaussian probability density function (pdf) is defined to get the degree of membership (association) of a training sample into a cluster (class). These membership values are used to define the global and class-wise mean training samples. Finally, the Gaussian probabilistic between-class and within-class scatter matrices are found separately in row and column directions of the image matrices. Two different Fisher´s criteria are defined based on these scatter matrices to generate the lower-dimensional discriminant features from the image matrices. Experiments on the AT&T (formerly ORL) and UMIST face databases show that the GPG-2DLDA method consistently improves the recognition rates in comparison with some subspace-based methods.