Fast Rotational Sparse Coding

We propose an algorithm for rotational sparse coding along with an efficient implementation using steerability. Sparse coding (also called \emph{dictionary learning}) is an important technique in image processing, useful in inverse problems, compression, and analysis; however, the usual formulation fails to capture an important aspect of the structure of images: images are formed from building blocks, e.g., edges, lines, or points, that appear at different locations, orientations, and scales. The sparse coding problem can be reformulated to explicitly account for these transforms, at the cost of increased computational cost. In this work, we propose an algorithm for a rotational version of sparse coding that is based on K-SVD with additional rotation operations. We then propose a method to accelerate these rotations by learning the dictionary in a steerable basis. We compare the proposed approach to standard sparse with experiments on patch coding and texture classification; in the latter case we report state-of-the-art results on the Outex_TC_00010 dataset.