Improving the fast back projection algorithm through massive parallelizations

Abstract

With sensor technologies rapidly improving, the need to process increasingly larger data sets is becoming the main bottleneck in many real time applications associated with persistence surveillance such as VideoSAR and volumetric SAR imaging. In many instances, the image fidelity is of utmost importance which can have implications when choosing the appropriate algorithm to generate the desired data products. The performance improvements afforded by algorithms such as the fast back projection (FBP) algorithm prove attractive for such environments. Unfortunately, even though the FBP algorithm is magnitudes faster than a traditional back projection algorithm it is still incapable of meeting the strict requirements of some of the aforementioned real time applications. However, the emergence of general purpose graphical processing units (GPGPUs) in recent years have afforded many scientific fields orders of magnitudes improvement in performance for a large variety of applications. This is also the case for the FBP algorithm. By distributing the processing across 480 processing cores located on a single video card, it possible to achieve substantial performance improvements compared to the serial FBP algorithm. Considering that many PCs are capable of housing three to four video cards, it is possible to obtain more than two orders of magnitude improvement in performance with the parallel approach. This technology provides the ability to process enormous datasets in the field without the need of supercomputers that have to date been the only means of keeping pace with the incoming data.