Multiscale image processing on the sphere

Abstract

We present linear filters for image processing in the case that the image data is given on the sphere rather than on a plane. Such spherical images occur in various situations in computer vision and computer graphics. The class of filters we present is derived from the spherical Gaussian kernel defined as the Green's function of the spherical diffusion equation. The derived filters include Laplacian of Gaussian, directional Gaussian derivatives, and their Hilbert transform. All computations are directly performed on the sphere without ever switching to a planar domain. These filters allow spherical image processing on multiple scales. We present results on images obtained from an omnidirectional camera. © Springer-Verlag Berlin Heidelberg 2002.