Bit-ordered tree classifiers for SAR target classification

Template-matching is the least sophisticated and most compute-intensive part of automatic target recognition (ATR) processing for synthetic aperture radar (SAR) applications. Complexity considerations dictate the use of low template densities in target signature space, while the extreme sensitivity of correlation processing to pose mandates the averaging of templates over a range of pose angles to achieve some generalization. Thus, the templates considered in ATR systems are often generated by noncoherently averaging target templates within a five degree pose-angle window, resulting in poor correlation gain. In this paper, we propose to dramatically increase the computational efficiency of correlation-based reasoning, using a completely different paradigm-the bit-ordered tree classifier (BOTC)-to enable high-density, high-confidence matching. Instead of performing, for example, 8-bit correlation between template and test images and comparing to a threshold, the BOTC makes selected binary comparisons to reach the same acceptance/rejection decisions with comparable operational characteristics, using far fewer computations. We report up to a two orders of magnitude speedup, compared to 8-bit correlation in preliminary testing on SAR target data from the MSTAR collection. We also investigate the efficient mapping of our novel BOTC technique to adaptive computing platforms such as field programmable gate arrays (FPGAs).