MODELING AND SEGMENTATION OF SPECKLED IMAGES USING COMPLEX DATA

This paper presents stochastic models and segmentation algorithms for speckled images, such as Synthetic Aperture Radar (SAR) images. The stochastic model developed is a two-level hierarchical random field model which consists of, at the higher level, a Gibbs random field governing the grouping of image pixels into regions and, at the lower level, speckle processes representing observations in the different regions, which are also modeled as random fields. In accordance with the physical phenomena that cause speckle, the single-look complex speckle process is modeled as a circularly symmetric complex Gaussian random field. With the assumption of a separable autocovariance for the complex Gaussian random field, the statistical description of the complex speckle becomes complete. Starting from the model for the single-look complex speckle process, different versions of the model are developed for multilook complex and single- and multilook intensity speckled images. In each case, the resulting hierarchical image model satisfactorily represents the spatial continuity between the regions of the image as well as the speckle correlation within each region. Maximum a posteriori segmentation algorithms using simulated annealing are developed for each of the models corresponding to the single- and multilook, complex and intensity speckled images. Parameters of the speckle processes are estimated as part of the segmentation algorithm. The segmentation algorithms are implemented on a wide class of synthetically generated and actual speckled images. The major finding implied by the experimental results is that segmentations using the complex amplitude data are significantly better than those using the intensity data, thus implying that, despite its noise-like appearance, the phase part of the data also carries useful information. © 1990 IEEE