Hyper-Parameter Selection on Convolutional Dictionary Learning Through Local ℓ0,∞ Norm

Abstract

Convolutional dictionary learning (CDL) is a widely used technique in many applications on the signal/image processing and computer vision fields. While many algorithms have been proposed in order to improve the computational run-time performance during the training process, a thorough analysis regarding the direct relationship between the reconstruction performance and the dictionary features (hyper-parameters), such as the filter size and filter bank’s cardinality, has not yet been presented.As arbitrarily configured dictionaries do not necessarily guarantee the best possible results during the test process, a correct selection of the hyper-parameters would be very favorable in the training and testing stages. In this context, this works aims to provide an empirical support for the choice of hyper-parameters when learning convolutional dictionaries. We perform a careful analysis of the effect of varying the dictionary’s hyper-parameters through a denoising task. Furthermore, we employ a recently proposed local $\ell_{0, \infty}$ norm as a sparsity measure in order to explore possible correlations between the sparsity induced by the learned filter bank and the reconstruction quality at test stage.