Texture analysis of SAR images applied to speckle filtering

This paper proposes a new model for the two-point statistics of spatial texture in SAR images. The autocovariance function is locally approximated by a 2-D anisotropic Gaussian kernel (AGK) in order to characterise texture by its local orientation and anisotropy. The estimation of texture parameters at a given scale is based on the gradient structure tensor (GST) operator and does not require the explicit computation of the autocovariance. Finally, a new filter called AGK-MMSE that takes into account this spatial information is introduced and compared to the refined MMSE filter. The proposed filter shows better performances in terms of texture preservation and structure enhancement