Abstract

Binary phase-only filters (BPOF) have attracted considerable interest because of their high diffraction efficiency and excellent correlation performance. When encoded with conventional methods for pattern classification, classification performance is not good because of the inherent disadvantages of matched filtering which is too sensitive to intraclass variation and too insensitive to interclass variation. In this paper we encode a BPOF with a simulated annealing (SA) algorithm to classify two similar character groups, the character P and the character R in several fonts. The SA algorithm chooses a binary state for each pixel in the BPOF through use of an energy function. This corresponds to supervised learning. During the training stage of the BPOF with the SA algorithm, five training characters in different fonts from each group are used. After training, the final BPOF is tested with five characters from each group that are in fonts different from the training characters. Zero classification error resulted. The computational requirement with the SA training algorithm is not excessive.

© 1989 Optical Society of America