Empirical Bayes least squares estimation without an explicit prior

M Raphan and E P Simoncelli

Published in SIAM Conf. on Imaging Science, May 2006.

This paper has been superseded by:
Least squares estimation without priors or supervision
M Raphan and E P Simoncelli.
Neural Computation, vol.23(2), pp. 374--420, Feb 2011.


Bayesian estimators often use a prior probability model, which is difficult to infer from noisy measurements. For additive Gaussian noise, however, the Bayesian least-squares estimator can be constructed directly from the logarithmic derivative of the noisy data distribution. We develop a local, adaptive approximation of this estimator, and use simulations to illustrate its behavior on various distributions. Despite its generality, the estimator performs well in denoising photographic images, compared with fitting a parametric prior.
  • Listing of all publications