A local spectral distribution approach to face recognition

This work presents a novel face recognition approach that utilizes the whole manifold structure of the face’s local spectral distribution. Local spectral features are extracted using Gabor wavelets, encoding at every image pixel the visual appearance of the surrounding patch by the vector of filter responses. The above representation provides a robust and discriminative statistical image description in the spatial frequency domain transform space. Parameterized by angle and scale, the manifold structure of the produced multidimensional point set contains both local and holistic information about the face image. In order to reduce redundancy and code efficiently the formed multivariate distribution, a neural vector quantizer is employed. The ensemble of the selected code vectors constitutes the spectral signature of a face image in the high-dimensional face space. The similarity between two face images is assessed by comparing the corresponding representative samples of the two distributions directly in the frequency space using the multivariate Wald–Wolfowitz test, a non-parametric statistical test dealing with the multivariate “Two-Sample Problem”. Its operation is based on the construction of the minimal spanning tree, which is an effective tool for preserving and utilizing the manifold structure of the data set. The new representation is both holistic, considering the features’ distribution as a whole, while at the same time utilizes local information extraction. Experimental results on four benchmark face databases demonstrate the favorable properties of the proposed methodology over traditional approaches particularly in the “single image case”.

► The presented statistical treatment tackles inherent facial noise or variability.
► No requirement for preprocessing or extraction of distinctive spatial features.
► The extracted distributions embrace local and global face information.
► The method is particularly suited for single face-image applications.
► The method deals successfully with the challenging small sample size problem.