|
|
|
“Computational Vision Algorithms for Synthetic Aperture Radar Imagery”
by Robert T. Frankot
October 1987
Two classes of models, computational vision models and stochastic models, are examined for synthetic aperture radar (SAR) images of natural terrain. Algorithms are developed for surface topography estimation, image registration, and texture synthesis.
Shape from shading techniques are used for extracting topographic information. Previous numerical solutions to the shape from shading problem estimated the surface derivatives without ensuring that they are integrable, a serious drawback. The performance of a previously developed shape from shading technique is substantially improved using a fast least-squares algorithm to enforce integrability. The resulting algorithm is then applied to SAR by representing the terrain surface height relative to the "slant plane" (a plane parallel to the line-of-sight) and accounting for the radiometric properties of SAR imagery.
For noisy imagery, such as SAR, low frequency surface information is difficult to recover from a single image. A fast Fourier transform implementation of the integrability projection provides an efficient method for combining low frequency surface information with the shading information. This technique may be suitable for combining the SAR imagery and the low resolution altimetry provided by Magellan to construct high resolution topographic maps of Venus. The resulting algorithm is applied to SIR-B SAR imagery and the surface reconstructions are compared with stereoscopically derived digital terrain models. The use of auxiliary low frequency information is tested, allowing estimation of reflectance map parameters and providing coarse surface structure to complement the surface details obtained from shading information. This simulates the Magellan scenario.
An automatic registration algorithm is used for matching image intensity predictions with the observed images. This registration algorithm matches two SAR images made from nearly orthogonal flight paths and matches a SAR image with an aerial photograph without detailed a priori knowledge of the terrain, two very difficult problems for images of hilly terrain.
Stochastic models for SAR image intensity based on nonlinear transformations of Gaussian random fields are introduced. Methods for selecting transformations to normality and model order are presented and tested on SAR imagery and synthesis of textures appearing in SAR imagery is demonstrated.