The combination of the IRAS focal plane arrangement, detector size, and scan pattern optimized detection of point sources in areas of the sky where the separation between sources was large compared to the sizes of the detectors. However it complicated the construction of images of regions containing spatial structure on the scale of arcminutes.
The desire for higher spatial resolution combined with the paucity of new infrared missions has inspired many efforts to extract high spatial resolution information from the IRAS data. The most widely accessible to the science community are the HIRES images distributed by the Infrared Processing and Analysis Center (IPAC), which are based on a Richardson-Lucy type of algorithm.
The HIRES process is very demanding computationally. A 1 degree by 1 degree field of typical scan coverage takes 1--2 hours of CPU time, for all 4 wavelength bands and 20 iterations. This motivated the port of the HIRES program to the massively parallel supercomputers currently available at Caltech. The new 512-node Intel Paragon Model L38 has a peak speed of 38.4 GFLOPS, 16 Gigabytes of memory, and 14 RAIDs that control 67.2 Gigabytes of disk.
With 8 nodes on the Paragon, a speed improvement of 5 times is achieved for a 1 degree by 1 degree field. Therefore the entire Paragon can be used to simultaneously process 64 square degrees, with an effective speedup factor of 320.
We also developed new algorithms which can effectively eliminate the most prominent artifacts in the HIRES images, namely cross-scan striping and ringing around bright point sources.
Cross-scan striping is caused by imperfect calibration of detector offset and gain. These are estimated within the HIRES scheme and compensated for. The resulting constrained algorithm decreases the striping power monotonically to a negligible level after about 10 iterations, as shown by Fourier analysis of the output image, unlike the unconstrained case, where striping builds up along with the resolution enhancement.
The ringing problem is tackled by incorporating an entropy prior to the goodness-of-fit function. This gives an image update rule which boosts the correction factors for bright pixels, thus extrapolating the high spatial frequency components and effectively suppressing the ringing artifact.
The parallelization and artifact reduction effort are complementary to each other. The artifact reduced algorithm allows the HIRES iterations to be carried out much further than before, achieving higher resolution without the contamination from artifacts. This in turn requires more computation time, which is provided by the parallel facilities.
We are therefore able to undertake an ambitious project --- HIRES mapping of the Galactic plane at 60 and 100 microns. These maps will represent a 20-fold improvement in areal information content over current IRAS maps and will be valuable for a wide range of scientific studies.