Joint Low Mutual and Average Coherence Dictionary Learning

Dictionary learning (DL) has found many applications in sparse approximation problems. Two important properties of a dictionary are maximum and average coherence (cross- correlation) between its atoms. Many algorithms have been presented to take into account the coherence between atoms during dictionary learning. Some of them mainly reduce the maximum (mutual) coherence whereas some other algorithms decrease the average coherence. In this paper, we propose a method to jointly reduce the maximum and average correlations between different atoms. This is done by making a balance between reducing the maximum and average co- herences. Experimental results demonstrate that the proposed approach reduce the mutual and average coherence of the dictionary better than existing algorithms.