On Encoding Neural Networks to Estimate the Atmospheric Point Spread Function in a Parallel Environment

Abstract

Images of the Earth's surface acquired by high-altitude aircraft or satellites are degraded by the intervening atmosphere. The imaging instrument records not only the signal of the targeted viewing area but also the radiance scattered into the field of view in the near by area. This effect can be characterized by an atmospheric point spread function (PSF). There are many parameters that may affect the PSF. To restore noisy-blurred images, one must understand which parameters influence the PSF and to what degree. This is very important for scientific applications that seek to extract information about environmental systems. In this paper, a design and implementation of a distributed representation scheme and neural networks are presented in order to estimate the atmospheric PSF. The representation scheme exemplifies the conjunctive coding and coarse coding techniques. Neural networks trained using such an appropriately structured representation generate a desired approximation of the PSF with satisfactory processing time. The system is implemented on a parallel system formed of personal computers and workstations. The maximum speed-up is achieved.