This document decribes the Spectrally Sampled Structural Subspace Features algorithm, which is abbreviated as 4SF. 4SF uses multiple discriminative subspaces to perform recognition. After geometric normalization of a face image using the automatically detected eye coordinates, illumination correction is performed using a local contrast enhancement algorithm. Face images are then represented using histograms of local binary patterns at densely sampled face patches. For each face patch, principal component analysis (PCA) is performed so that $98.0\%$ of the variance is retained. Given a training set of subjects, multiple stages of weighted random sampling is performed, where the spectral densities (i.e. the eigenvalues) are used for weighting. For each stage of random sampling, linear discriminant analysis (LDA) is performed on the randomly sampled components. The LDA subspaces are learned using subjects randomly sampled from the training set (i.e. bagging). Finally, distance-based recognition is performed by projecting the LBP representation of face images into the per-patch PCA subspaces, and then into each of the LDA subspaces learned. The sum of the Euclidean distance in each subspace is the dissimilarity between two face images.
You are granted permission for the non-commercial reproduction, distribution, display, and performance of this technical report in any format.