Family of heterocorrelation filters

Abstract

We introduce a new concept in pattern recognition that we call heterocorrelation. Contrary to standard approaches, heterocorrelation allows correlation of different images solely by modifying the filter's intensity transmissivity in the areas of greatest phase mismatch relative to the phase of the stored template. This approach can convert a single object recognition filter to a multiple object classification filter. We develop three algorithms and successfully test each of them for three completely different input cases: one that uses geometric input images with very little common edge information, one that uses these same images encoded with high spatial frequency information, and one that uses synthetic aperture radar (SAR) images that have almost no edge information. The third algorithm provides shift-invariant heterocorrelation with equalized in-class correlation peaks, and eliminates the possibility of any higher-order side peaks for the added in-class objects. We also demonstrate how the use of a heterocorrelation filter can improve the performance of conventional multiplexed filters, such as synthetic discriminant filters, as well as increase their template storage.