Iterative Dictionary Learning Based on Nonconvex Sparse Regularization and Adaptive Parameter Control

摘要

In this paper, we study the dictionary learning algorithms for sparse representation with nonconvex sparse regularization. By splitting and recombination technics, we derive out the updating formula for sparse coefficients in block wise based on coordinate descent method. The update formula for atom is obtained by minimizing a convex least square problem. In addition, we introduce a method to adaptively control the penalty parameter based on a tradeoff between the sparsity and error. Simulation and application experiments are carried out and experiment results show the proposed algorithm is much more accurate in extraction of the transient impulsive signal than the Ll norm based counterpart and its running time is much less than that of the K-SVD algorithm.