A Spectral Graph Theoretical Approach to Oriented Energy Features

In this work, we propose a novel method for determining oriented energy features of an image. Oriented energy features, useful for many machine vision applications like contour detection, texture segmentation and motion analysis, are determined from the filters whose outputs are enhanced at the edges of the image at a given orientation. We use the eigenvectors and eigenvalues of graph Laplacian for determining the oriented energy features of an image. Our method is based on spectral graph theoretical approach in which a graph is assigned complex-valued edge weights whose phases encode orientation information. These edge weights give rise to a complex-valued Hermitian Laplacian whose spectrum enables us to extract oriented energy features of the image. We perform a set of numerical experiments to determine the efficiency and characteristics of the proposed method. In addition, we apply our feature extraction method to texture segmentation problem. We do this in comparison with other known methods, and show that our method performs better for various test textures.