Parallel Image Processing Applied to Radar Shape-from-Shading

Various widely used radar image processing algorithms require considerable computing resources but can take advantage of a parallel implementation. We focus on the Shape-from-Shading (SfS) algorithm in its application to radar images. A given serial version of the SfS algorithm was parallelized and improved to handle large images. We experiment with parallelization techniques such as data decomposition, the manager/worker method and dynamic load balancing with double buffering. The parallel version of SfS was ported to two supercomputers: Meiko’s CS-2HA and Intel’s PARAGON XP/S-A4 distributed memory machines, and to a Cluster of Workstations (CoW) made up of SGI’s Indies. Important results concerning the performance of the parallel SfS implementation on those architectures are presented and compared to each other, showing that 14 processors can speed up SfS by up to 13 times over the use of a single processor.