Fast rotational dictionary learning using steerability
Mike McCann, EPFL STI LIB

Meeting • 08 May 2018

Abstract
In this talk, I will present work I have done with Adrien on rotational (also called "rotation-invariant" or "rotational-equivariant") sparse dictionary learning (DL). Starting with a set of training data, the goal of DL is to find a dictionary comprised of elements (called "atoms"), such that each element of the training data can be well-approximated by a linear combination of a small number of atoms. In image processing applications, these linear dictionaries struggle to capture the rotational and translational relationships between patches, resulting in dictionaries with low approximation power and a lack of specificity. These problems can be addressed by explicitly accounting for these transformations in the problem formulation, but the resulting learning algorithms are usually impractically slow. Here, we present a new technique for fast rotational DL which uses a discrete steerable basis to accelerate the learning. We demonstrate the usefulness of the technique in both coding and texture classification.