Neural networks for extraction of weak targets in high clutter environments

Because of the statistical nature of many types of clutter, a detection device must set a high threshold in order to maintain a reasonable false-alarm rate. However, by selecting this threshold setting, detections of small and medium size targets can be missed. An old but previously impractical technique for improving performance was to use all contacts from several scans and employ a very large bank of matched filters. This could achieve a detection on one or more of all possible target trajectories formed from stored contacts for each input detection. Neural network hardware offers new opportunities to implement such techniques. It is shown that feedforward and graded-response Hopfield neural networks can implement the optimum postdetection target track receiver. For the Hopfield net, the spurious states correspond to the important case of multiple track detection. Finally, the author presents simulations that show that substantial signal-to-noise gain can be achieved.<>