A robust method for parameter estimation of signal-dependent noise models in digital images
Bruno Aiazzi, Luciano Alparone, Stefano Baronti · 2002
A class of signal-dependent noise models is discussed, with reference to the cases of film-grain and speckle noise, which are commonly encountered in image processing applications. The model is uniquely defined by the variance of the zero-mean random noise (independent of the signal) and by the gamma exponent which rules the dependence with the signal. A robust procedure for measuring such parameters directly from the noisy images is presented. First, the gamma coefficient is estimated from at least three homogeneous non-textured regions. Then, the noise variance is determined as the mode of the histogram of the ratio between the local variance, and the local mean raised to twice the gamma. Computer simulations show the high accuracy of the results.