<title>Maximum likelihood technique for blind noise estimation</title>

Abstract

We propose a novel technique for estimation of image noise amplitude without a priori signal information. Knowledge of the normalized noise distribution is used to construct an approximate Wiener filter parametrized by the estimated noise amplitude. For a given noise amplitude, the resulting signal estimate is subtracted from the image to produce a sample noise estimate. The estimated noise amplitude is varied in order to maximize the probability that the noise estimate is a sample of the known noise distribution with the estimated variance. Probability is measured by the (chi) ² distribution. The technique is tested for several images by adding stationary zero-mean Gaussian noise with varying amplitude. The variation of estimated versus added noise variance is very nearly linear with unit slope for all of the images tested. The estimated noise variance for images with no added noise is generally small compared to the signal power unless the signal power spectrum is nearly white.