Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!mit-eddie!apollo!johnf
From: johnf@apollo.COM (John Francis)
Newsgroups: comp.graphics
Subject: Re: Color Quantization
Message-ID: <4388acb6.7f32@apollo.COM>
Date: 30 May 89 16:47:00 GMT
References: <310@celit.UUCP> <311@celit.UUCP>
Reply-To: johnf@apollo.COM (John Francis)
Organization: Apollo Computer, Chelmsford, MA
Lines: 14

I have implemented Paul Heckbert's Color Quantization algorithm, but
with the following modification (suggested in the original paper):

Choose the box to split (and the axis along which it is to be split)
such that at each step you get the largest reduction in the variance
(this is equivalent to at each step choosing the split that reduces
the overall root mean square error by the largest amount).
I find that this gives a better result than Median Cut (the original
method) or Mean Cut.
I have also experimented with giving different weights to the three
RGB axes. Weights of R=0.299, G=0.587, & B=0.114 produce a slightly
different result in the overall appearance of the resulting picture
which most of the people I have asked seem to feel is a very slight
improvement.  Your mileage may differ.