"Optimal" imaging with curvelets

In this paper we present a non-linear edge-preserving solution to linear inverse scattering problems based on optimal basis-function decompositions. Optimality of the basis functions allow us to (i) reduce the dimensionality of the inverse problem; (ii) devise non-linear thresholding operators that approximate minimax (minimize the maximal mean square error given the worst possible prior) and that significantly improve the signal-to-noise ratio on the image. We present a reformulation of the standard generalized least-squares formulation of the seismic inversion problem into a formulation based on thresholding, where the singular values, vectors and linear estimators are replaced by quasi-singular values, basis-functions and thresholding. To limit the computational burden we use a Monte-Carlo sampling method to compute the quasi-singular values. With the proposed method, we aim to significantly improve the signal-to-noise ratio (SNR) on the model space and hence the resolution of the seismic image. While classical Tikhonov-regularized methods only gain the square-root of the SNR on the data for the SNR on the model our method scales almost linearly. This significant improvement of the SNR allows us to discern events at high frequencies which would normally be in the noise.

https://slim.gatech.edu/Publications/Public/Conferences/SEG/2003/Herrman...