Re: Fuzzy Pattern Recognition

WSiler (WSiler@aol.com)
Mon, 18 May 1998 02:40:04 +0200 (MET DST)

In a message dated 98-05-01 05:43:17 EDT, you note that the size of a fuzzy
rule base tends to increase with the number of input variables, and ask how
this can be dealt with in fuzzy pattern recognition. I can recall two possible
approaches.

1) William Combs of Boeing in Seattle, USA has an approach which enables the
number of rules to increase linearly with the number of input variables rather
than as the square.

2) We take a different approach. One can consider the number of input
variables as the number of pixels in an image. Not only would this give a
ridiculous number of rules, but the rules would be quite difficult to write.
We pursue the classical three-step approach: phase 1, segmentation of the
image into regions of interest; phase 2, feature extraction on these regions;
phase three, classification of the regions. We employ primarily conventional C
programs on the first two phases, although fuzzy AI may be helpful as an aid
to segmentation. Feature extraction is always done by straightforward C
programs. This reduces the input variables to the fuzzy rule-based
classification program to the number of features, with probably several
instances of the classification rules being made concurrently fireable because
of concurrent input of a number of regions. This makes for a manageable number
of rules.

William Siler