Multiscale-combined inversion of convolutional model based on orthogonal wavelet transform

JW Ma,HZ Yang,YP Zhu

DOI: https://doi.org/10.1142/9789812777362_0045

2002-01-01

Abstract:Traditional techniques of multiscale inversion pay a lot attention to the truth that the local minima at coarser scale are far apart from each other, thus provide opportunities for finding the global minimum efficiently at the coarser scale and then switch to the finer scales. But they have not taken full advantage of the coherent relation between scales. In this paper, a method of multiscale-combined inversion (MCI) is first proposed to exploit the coherent relation between scales, compared to the traditional techniques, which we prefer to call multiscale-independent inversion (MII). Since the model (or signal) at coarser scale is a "smoother" version of the model (or signal) at finer scale, inversion result derived from the coarser scale may be considered as certain kind of constraints in the following scales, not merely as a good initial guess for the next scale. MCI is a promising method because of some advantages such as little dependence on initial model, efficiency of convergence and robustness. As one realization of MCI, a new formulation of scale-constrained Least Mean Square (LMS) method has been constructed, which we point out is a certain kind of generalizations of the traditional finite power constrained LMS. Several numerical results show the effectiveness and potential of the method.