Automatic Target Recognition (ATR) ATR: Background Statistics and The Detectionof Targets in Clutter

摘要

This research investigated signal processing of two dimensional signals for thedetection of targets in noise, particularly in complex background pattern noise. The researchers hypothesized that this type of noise was vulnerable to non-linear processing. They investigated whether the human eye/brain acting as a surrogate for a non-linear processor could outperform an optimum linear processor in a quantitative sense. The researchers did this by conducting computer experiments to determine the ability of an operator and an optimum linear filter to determine a known pattern's presence or absence in a noisy image. The performance of both the operator and optimum linear filter are recorded as probability of detection, probability of false alarm pairs, which the researchers use to determine effective signal-to-noise ratio. The performance of man versus machine (optimum linear filter) is compared quantitatively using the effective signal-to-noise ratio. Operator and machine/filter are tested against circular targets in Random White Gaussian noise and in satellite images. The researchers report that the machine filter outperforms the man when the details of both target and background are known in advance, but the man outperforms the machine/filter when the details are known only in a statistical sense.