Structured Dictionary Learning Based on Composite Absolute Penalties

In this paper, we focus on the problem of learning dictionaries with structural features for the sparse representations of natural images. Dictionaries learned by traditional techniques such as MOD, K-SVD lack structure. Each atom of them is treated independently and the possible relationships are not fully explored, which is insufficient for some cases. We propose a framework for structured dictionary learning by integrating the Composite Absolute Penalties (CAP) into the K-SVD algorithm. Atoms of the learned dictionary are laid out in a predefined fashion, i.e., group or tree structure. Such a setting is more appropriate to exploit the latent relationships existing between the patches of natural images. Experiments show that dictionaries learned by our method achieve better results for image restoration tasks. Our approach can also be integrated into other sparse representation-based applications of image processing.