Despeckling Synthetic Aperture Radar images with cloud computing using graphics processing units

This paper presents the implementation of Synthetic Aperture Radar (SAR) image enhancement and information extraction techniques using multicore Graphic Processing Units (GPUs) connected into a cloud computing environment. The Bayesian approach to SAR image despeckling and information extraction is presented. The first order Bayesian inference is used to estimate a maximum a posteriori (MAP) estimate. A prior is modeled using Gauss-Markov Random Fields (GMRF). The second order Bayesian inference is used to find the best model parameters, which represent texture information in SAR images. The algorithm is rewritten in matrix form to fully exploit GPUs computing power. GMRF on GPU give good results for despeckling and information extraction, but this algorithm is also very computationally demanding. This paper presents the implementation of the MAP estimator and evidence maximization using GPUs. The image is divided into subblocks and each subblock has as many threads as there are pixels inside the subblock. The estimation of MAP solution and its corresponding model is computed separately in each thread in the sub-block The GPUs implementation of the model-based despeckling is about 95-times faster as the implementation of the same algorithm on multicore CPUs on a development system. The whole cloud computing environment has proven to be fast and robust.