Economic descriptions of objects are essential for model-based ATR. We develop economic target descriptions based on high range resolution target returns, utilizing wavelet multiresolution representations and Tree Structured Vector Quantization, in its clustering mode. The algorithm automatically constructs the multi-scale aspect graph of the target, which is a most efficient model that guides the on-line ATR algorithm. This results in a progressive coding of the target model information and in an extremely efficient, hierarchical indexing of the stored target models. As a final outcome we obtain extremely fast recovery, search and matching during the on-line ATR operation. In addition, these scale space (or multiresolution) representations can be used to generate target fingerprints across a range of scales, unavailable heretofore, which aid substantially in target recognition even with occluded or spurious data.
We have Successfully addressed the problem of reducing the target model representations with respect to viewpoint variations and other sensor parametric variations. Our method can be viewed as a quantization of the space of sensing operations. The resulting aspect graph is a (relational) graph representation of this quantization. Aspect graphs of target radar returns are generated algorithmically. Since our off-line model/parameter tuning methods are based on general vector quantization, our methods extend naturally and efficiently to multi-sensor data: LADAR, TV, mmWave, SAR, etc..
We also investigate the so-called new target insertion problem in a fielded ATR system, and the required fast re-programmability of the ATR system. We compare the performance and cost (both computational and hardware) of ATR algorithms based on the parallel use of single target aspect graphs vs ATR algorithms using the combined aspect graph for the group of targets under consideration. We show that efficient real-time ATR algorithms can be constructed using the aspect graph of each target in a parallel computation. The resulting architecture includes wavelet preprocessing with neural networks postprocessing. We use synthetic radar returns from ships as the experimental data to demonstrate the performance of the resulting ATR algorithm.