Abstract

The best threshold level to obtain a given performance for a matched filter in pattern recognition may depend on the noise level. If the noise level is unknown, varies over the image, or is signal-dependent, the performance of the matched filter may be severely affected. This is particularly true of composite filters, which are linear combinations of matched filters used for pattern recognition or classification. This effect may be corrected for by adding an extra noise-correction term to the matched filter. Experimental results show improved performance of composite filters when the extra term is included in the filter.

© 1986 Optical Society of America