Fast Rotational Sparse Coding
This work addresses a domain-specific problem in image processing for tasks like inverse problems and compression, but it is incremental as it builds on existing K-SVD methods with rotational extensions.
The authors tackled the problem of sparse coding's inability to capture image structures like edges at different orientations by proposing a rotational sparse coding algorithm based on K-SVD with steerable basis acceleration. The result is a fast algorithm that performs comparably to standard sparse coding in patch coding and texture classification experiments.
We propose an algorithm for rotational sparse coding along with an efficient implementation using steerability. Sparse coding (also called 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 computation. 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. Our experiments on patch coding and texture classification demonstrate that the proposed algorithm is fast enough for practical use and compares favorably to standard sparse coding.