Abstract
We demonstrate an adaptive multi-layer optical classifier performing
single pulse radar target recognition to identify isolated aircraft
targets with varying orientation and/or range from the radar. The
system uses optically-calculated time-frequency representations as its
internal representation, and in particular the triple autocorrelation
[1] due to the natual range invariance of this feature. This approach
increases the separability of the input data by nonlinearly mapping it
into a higher dimensional feature space. Serial processing of the
optically computed feature vector using CCD detectors and electronic
postprocessing restricts system throughput since the massive
quantities of data overburdens electronic digital postprocessing,
whereas adaptive optical classification avoids this electronic
bottleneck. We have previously demonstrated a broadband communications
signal classifier using a non-adaptive classifer [2]. In this paper we
report an adaptive multi-layer optical classifier and present
experimental classification results in which the neural layer learns
to identify optically computed triple autocorrelation representations
of a training set of aircraft radar range profiles. The generalization
performance to untrained low resolution profiles was excellent.