De-speckling of SAR imagery based on multi-resolution analysis: a comparison of ML and MAP estimators

In this paper, Generalised Gaussian PDFs are tailored to wavelet detail coefficients of both reflectivity and speckled intensity images. A MAP solution is derived from statistics calculated in the image domain by exploiting the fact that undecimated wavelet coefficients are given by convolving the image with a linear shift-invariant equivalent filter. Extensive experiments and comparisons demonstrate that, though theoretically and procedurally different, the ML and MAP undecimated wavelet-domain estimators yield comparable ENIL on simulated data, with a computational complexity of the MAP approach several times greater than that of ML filtering. However, on true SAR images the novel MAP filter visually outperforms the former ML filter, specifically concerning texture restoration and preservation of point targets.