Directional polynomial wavelets on spheres

TL;DR

This paper develops discrete tight frames of polynomial wavelets on spheres that are direction-sensitive, steerable, and localized, enabling efficient analysis of anisotropic features like edges with manageable sampling requirements.

Contribution

It introduces a new construction of discrete tight frames of polynomial wavelets on spheres that are directionally adjustable and suitable for anisotropic feature analysis.

Findings

Wavelets are steerable and highly localized in space.

Sampling complexity is comparable to isotropic cases due to limited directionality.

Wavelet expansions converge rapidly in $L^p$ spaces.

Abstract

In this article, we construct discrete tight frames for $L^2(\mathbb{S}^{d-1})$, $d\geq3$, which consist of localized polynomial wavelets with adjustable degrees of directionality. In contrast to the well studied isotropic case, these systems are well suited for the direction sensitive analysis of anisotropic features such as edges. The price paid for this is the fact that at each scale the wavelet transform lives on the rotation group $SO(d)$, and not on $\mathbb{S}^{d-1}$ as in the zonal setting. Thus, the standard approach of building discrete frames by sampling the continuous wavelet transform requires a significantly larger amount of sample points. However, by keeping the directionality limited, this number can be greatly reduced to the point where it is comparable to the number of samples needed in the isotropic case. Moreover, the limited directionality is reflected in the wavelets being steerable and their great localization in space leads to a fast convergence of the wavelet expansion in the spaces $L^p(\mathbb{S}^{d-1})$, $1\leq p \leq \infty$.