Fractional Two-Dimensional Linear Discriminant Analysis

Two-dimensional linear discriminant analysis (2DLDA) is a dimensional reduction technique widely applied to image recognition, particularly to the task of human facial recognition. Despite this, one of the potential limitations of the standard 2DLDA is its inability to model a discriminative feature subspace with the appearance of undesirable variations (illumination changes, face expressions, etc.) in the original image samples. To address this limitation, fractional 2DLDA (F2DLDA) is suggested in this paper. F2DLDA uses fractional transformed images-as-matrices to generate a projection matrix through maximization of the fractional 2D Fisher’s criterion. In this manner, the constructed feature subspace by F2DLDA reflects the most discriminative information of the image samples and is comparatively less sensitive to undesirable variations. Results from a set of experiments affirm the viability of the proposed F2DLDA in face recognition.