SAR speckle mitigation by fusing statistical information from spatial and wavelet domains

Abstract

We propose a novel algorithm for the de-speckling of SAR images which exploits a priori statistical information from both the spatial and wavelet domains. In the spatial domain, we apply the Method-of-Log-Cumulants (MoLC), which is based on Mellin transform, in order to locally estimate parameters corresponding to an assumed Generalized Gaussian Rayleigh (GGR) model for the image. We then compute classical cumulants for the image and speckle models and relate them into their wavelet domain counterparts. Using wavelet cumulants, we separately derive parameters corresponding to an assumed generalized Gaussian (GG) model for the image and noise wavelet coefficients. Finally, we feed the resulting parameters into a Bayesian maximum a priori (MAP) estimator, which is applied to the wavelet coefficients of the logtransformed SAR image. Our proposed method outperforms several recently proposed de-speckling techniques both visually and in terms of different objective measures. © Springer-Verlag Berlin Heidelberg 2007.