Space-based sensors currently transmit large amounts of raw data to their receiving stations, forcing large bandwidths to be used and large data archives maintained. One possible solution to this data management problem is the use of more intelligent processing onboard the spacecraft. Neural networks are proposed as a trainable, mutable means of achieving sensor signal processing, sorting, and classification before infrared (IR) focal plane data leaves the cryogenic seal. Two questions arise: (1) can a neural network of appropriate size learn a sensor processing classification problem, and (2) will there be devices to implement that neural network as a smart focal plane. In this paper we concentrate on the first issue and show successful results in simulation for a massively parallel, distributed neural network solution to the closely spaced object (CSO) recognition problem, using a defocused JR sensor model with noise, and an uncooperative object or CSO at large distance. The back-propagation learning method is used to train the network. Architectural issues regarding the construction of a prototype neuromorphic focal plane device are also discussed.

Date Published: 1 November 1990
PDF: 8 pages
Proc. SPIE 1339, Materials, Devices, Techniques, and Applications for Z-Plane Focal Plane Array Technology II, (1 November 1990); doi: 10.1117/12.23006

Published in SPIE Proceedings Vol. 1339:
Materials, Devices, Techniques, and Applications for Z-Plane Focal Plane Array Technology II
John C. Carson, Editor(s)