Local anisotropy analysis for non-smooth images

Identification of local anisotropy and determination of principal axes are addressed through different methods that are designed to be tolerant to the non-smooth character of images on the pixel scale. These different tools are validated on various examples, and their performances are compared. The most powerful, robust and accurate method consists in computing the curvature tensor of the autocorrelation function of regularized images using fast Fourier transforms. (c) 2005 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved.