Previous
abstract 
Next abstract
We have implemented on a variety of parallel processing computers the infrared cirrus filter developed by Appleton, Siqueira, and Basart (see the Proceedings on Astronomical Data Analysis Software and Systems I Conference, pg. 283). The filter uses the techniques of mathematical morphology (also known as morphological image processing). Initial results indicate a $\geq 15x$ reduction in the average level of cirrus emission in the $100\mu$ IRAS image of the M81/M82 field.
The key feature of morphological image processing is its ability to discriminate features based on shape (hence the name). The shape to be probed is contained in the structuring element. Initial results with the filter reported above used a gaussian structuring element.
The filter is computationally expensive (the initial implementation required 8 hours on a workstation), and so we were motivated to explore the use of parallel processors to reduce the filtering time. The parallel processing platforms we have used to date are the MasPar MP-1 and MP-2, the Intel Paragon, and the Kendall Square Research KSR-1. We report the performance attained with each platform, and compare to those of traditional vector processors such as a Cray Y-MP and Convex C3240, workstations such as a DECstation 5000, and the new generation super-workstations made by DEC and HP. For example, the MasPar MP-2 with 16K processors is 350x faster on this problem than a DECstation 5000/240, and is 10x the performance of a Y-MP processor.
The dramatically faster performance available from the Maspar has made it much easier to experiment with the parameters of the cirrus filter. We report the results of filter tests made using disk and gaussian structuring elements with a variety of amplitudes.
This work is funded by the NASA High Performance Computing and Communications Program (HPCCP).
