Next: L norm and non
Up: SYNTHETIC EXAMPLE
Previous: Predictive deconvolution
I tried the same processing with exponential and gaussian noise. Unfortunately,
the L1 minimization is no more efficient. The noise samples now have various
amplitudes, about the same repartitions as the original trace, and cannot be
isolated from the seismic samples by the original criterion of small
weight: they have all kinds of weighting. The L1 norm itself does not
improve the result of the L2 deconvolution; the most usual way to suppress
the noise is to use large damping factors, which is not a typical
characteristic of the L1 norm. Moreover, even if the noise n is
gaussian, white, and uncorrelated with the seismic data, the L1 normal
equations cannot use these properties, because the correlations terms (nTn,
yTn) will be weighted by the matrix W, and replaced by nTWn and yTWn
(which may not vanish). I will suggest in the last part a method to
cope simultaneously with two kinds of noise, for example gaussian and spiky
noise, using L1 and L2 norms simultaneously.
Next: L norm and non
Up: SYNTHETIC EXAMPLE
Previous: Predictive deconvolution
Stanford Exploration Project
1/13/1998