BLIND NON-WHITE NOISE REMOVAL IN IMAGES USING GAUSSIAN SCALE MIXTURES IN THE WAVELET DOMAIN
Javier Portilla · 2004
Gaussian scale mixtures (GSM) capture two basic properties of the wavelet coefcients responding to natural images, namely 1) high kurtosis marginals, and 2) positive covariance between neighbor coefcient amplitudes. These features are not shared by Gaussian or lower kurtosis noise sources. Therefore, GSM models provide a means to separate the noise from the signal for an observed corrupted image. A local model consisting of a GSM term plus Gaussian additive noise with arbitrary covariance is used to estimate rst the noise covariance at each wavelet subband, applying a generalized expectation maximization algorithm. Then the original wavelet coefcients are estimated from the noisy observations using an efcient Bayes Least Squares technique. Both steps are fully automatic.