Multi-scale kernel basis and iterative orthogonal matching pursuit for sparse approximation

Function basis and approximation algorithm are two key elements in sparse representation. In this paper, some cardinal spline kernel basis and translation invariant fast decreasing kernel basis are presented, an iterative orthogonal matching pursuit algorithm (IOMP) is proposed, which is based on iterative update of hermitian inverse matrix. Experiments and comparisons on sparse representation of signal and regression datasets demonstrate that the proposed multi-scale kernel basis and iterative orthogonal matching pursuit algorithm (IOMP) are good at fast computing sparse approximation.