Adaptive segmentation of speckled images using a hierarchical random field model

A two-level hierarchical random-field model developed for speckled images and, in particular, for synthetic-aperture radar (SAR) imagery, is described. At the higher level is a Gibbs random field governing the grouping of image pixels into regions, while at the lower level are speckle processes for the different regions, which are also modeled as random fields. In accordance with the physical phenomena that cause speckle, the speckle processes are modeled as circularly symmetric complex Gaussian random fields. As with real Gaussian fields, certain forms of autocovariance of complex Gaussian field lead to Markovianity properties. With the assumption of a separable autocovariance for the complex Gaussian random fields, local joint statistics of the resulting speckle-intensity fields (magnitude-squared of the complex field) are determined. The speckle model concurs with the known marginal statistics and also accounts for the autocorrelation observed in actual speckled images. A MAP segmentation using simulated annealing and based on the hierarchical model is presented. The algorithm is adaptive in that it recursively segments the image and estimates the model parameters necessary for the segmentation