Heuristic Based Learning of Parameters for Dictionaries in Sparse Representations

Abstract

Sparse representation has attracted attention recently by successful applications in the computer vision domain. The success of these methods depends on the learned dictionary as it represents the latent feature space of the data. Different parameters affect the dictionary learning process like the number of atoms and sparsity limit. Generally, these parameters are learned through trial and error experimentation which requires a lot of time. In the literature, no approach is seen that attempts to relate these dictionary parameters to the data. In this paper, we propose heuristics for this problem. These heuristics use statistical properties of the data to estimate dictionary parameters. The proposed heuristics are applied to several datasets.