Skewed log-stable model for natural images pixel block-variance

This work presents a log-stable model for natural images block-variance. Exponential and halfnormal distributions have been previously used to model block-variance, but they were employed to fit images for which the assumption of constant intra-block variance does not hold. We show that when this assumption holds, the log-stable model yields a much better fit in an ML sense. We use a computationally efficient method for estimating the log-stable parameters through the empirical Kullback-Leibler divergence, which is asymptotically optimum in an ML sense, and show the validity of the lognormal distribution as an approximation with closed-form formulas for the ML parameter estimation.