Improved Structured Dictionary Learning with Adaptive Block-size Using K-means Clustering

In recent years, a lot of attention has been given towards generating block sparse representation of the data. To derive the unknown block structure, the existing works employ the agglomerative clustering with known maximum block size as stopping criterion instead of any threshold. As a consequence of that, the number of atoms in each block cannot be optimized. Towards overcoming this challenge, we propose to use the well known k-means algorithm in the block formation. In this way, no constraint is imposed on the block size unlike done in the existing works. The performance evaluations have been done on the image super-resolution task using the Berkeley Segmentation Dataset. The approach is noted to yield significantly better peak signal-to-noise ratio than the existing block-structured dictionaries techniques.