A hierarchical, automated target recognition algorithm for a parallel analog processor

A hierarchical approach is described for an automated target recognition (ATR) system, VIGILANTE, that uses a massively parallel, analog processor (3DANN). The 3DANN processor is capable of performing 64 concurrent inner products of size 1/spl times/4096 every 250 nanoseconds. A complete 64/spl times/64 raster scan of a 256/spl times/256 image can be evaluated by the 3DANN with its 64 modifiable templates in about 16 milliseconds. To fully utilize the analog processor and accommodate its high bandwidth capabilities, the vectors (templates) loaded on the 3DANN perform dimensionality reduction for a backend set of classifiers. The templates used in this ATR algorithm are hierarchically generated sets of eigenvectors taken from a partitioned set of library object images. As information is accumulated about the target (e.g. object class), a more refined set of eigenvectors reflecting this knowledge can be loaded and more specialized classifiers utilized. The classifiers provide information related to the ATR task: location, class, sub-class, and orientation of target(s). We report some preliminary results that examine the performance of orientation classifiers. With no knowledge about object class or orientation, a neural network achieves 94.2% in determining to which one of three classes from vertical (30/spl deg/, 45/spl deg/, or 60/spl deg/) an object image is oriented (/spl plusmn/30/spl deg/). Using an eigenvector template set generated from a distribution where both object class and orientation are known, a neural network classifier achieves 96% in orienting the object to within /spl plusmn/22.5/spl deg/. This information can be used to load even more specific eigenvector sets which should lead to more accurate object location during tracking and an enhancement in object recognition tasks.