Path: utzoo!attcan!utgpu!utstat!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!ucbvax!pasteur!ames!eos!jbm
From: jbm@eos.UUCP (Jeffrey Mulligan)
Newsgroups: comp.graphics
Subject: Re: Color Quantization
Message-ID: <3800@eos.UUCP>
Date: 30 May 89 21:00:35 GMT
References: <8490@venera.isi.edu>
Organization: NASA Ames Research Center, California
Lines: 65

From article <8490@venera.isi.edu>, by raveling@venera.isi.edu (Paul Raveling):
> In article <9445@polya.Stanford.EDU> rokicki@polya.Stanford.EDU (Tomas G. Rokicki) writes:
>>
>>Wouldn't it make sense to do color selection in the HSV cone rather
>>than in the RGB cube, since the `distances' in HSV are more relevant
>>to perceived differences in color?
> 
> 	Maybe, but it's not clear.

	[ text deleted ]

> 	There are also some applications of quantization that would
> 	need an approach not based on human perception.  I'm thinking
> 	of things such as multispectral imaging, including infrared,
> 	UV, X-ray, and magnetic resonance imaging.  Some applications
> 	might call for a "linear" quantization, some might benefit
> 	by weighting for a "non-human" domain.

I'm not sure what you have in mind here.  With these new imaging
techniques, it would seem to me that you can divide the problem
into two parts:  first, what are the features that you want to
detect?  and second, how can you process the image to make those
features most visible to a human observer while minimizing the
noise or number of false targets?  The first part doesn't have
anything to do with vision, while the second part doesn't have
anything to do with anything except vision.  What is the "non-human"
domain you are thinking of?


There is an important point that has not been mentioned: namely that
the spatial character of the acceptable quantization errors is
different for luminance and chromatic errors.  The visual system
is blind to chromatic variations at medium to high spatial frequencies
at which luminance modulations can still be clearly resolved.
This is why it is generally acceptable that the chroma channels
on NTSC video signals have lower bandwidths than the luminance signal.

A few years ago I played around with a technique to try and exploit this.
Imagine dithering an achromatic (black/white) grating.  Since the
R, G and B signals are the same, the dithered R, G and B images
will be the same, so the quantization noise in the final composite
image will be pure luminance noise with no chromatic error.
Since the visual system is relatively insensitive to high frequency
chromatic variation, it seemed that it would be preferable to
introduce chromatic errors if it could be traded off to reduce
the (more visible) luminance errors.  The way I tried to do this
was by using a modified error diffusion algorithm:  the R, G and B
images were processed in parallel; after pixel each was quantized
the resulting error was partitioned into luminance and chromatic
components.  This error was then "diffused" with different spread
functions for each of the components, with the normal weights being
used for the luminance error, and larger spreads used for the two
chromatic components.  Before quantizing the next pixel, the error
signals were converted back to R, G, and B components.  This worked,
but not surprisingly was slower than the already slow normal error
diffusion.  I never worked out a theory of the relation between
the error diffusion spread function and the resulting noise spectrum,
but the results were qualitatively as expected.


-- 

	Jeff Mulligan (jbm@aurora.arc.nasa.gov)
	NASA/Ames Research Ctr., Mail Stop 239-3, Moffet Field CA, 94035
	(415) 694-6290