2006.12.05 15:52 "[Tiff] Grayscale, or is it?", by Joris Van Damme

2006.12.06 20:56 "Re: [Tiff] Re: Tiff Digest, Vol 31, Issue 3", by Joris Van Damme

Graeme,

Either we're interpreting "straight", "linear" and "gamma"

differently,

or perhaps you are confusing the processing applied, with the state of encoding?

That does indeed seem to be a common confusion in this thread.

The standard conversion from RGB to gray, documented in the textbooks, is grey level = .33R+.5G+.17B. So a unitary conversion back to RGB is fine.

And of course this equation is acknowledged to be wrong - you can't combine gamma encoded values like this to give a correct greyscale. To do it properly, you have to turn the values into linear light encoding, then weighted sum them, and perhaps re-encode with gamma again. The NTSC system adopted the above simply for expedience - it's tricky to decode into linear light and back again using vacuum tubes - so they compromised on the monochrome reproduction.

It's totally off-topic, but interesting nonetheless, so here's my two cents...

Things depend on your colour model, or at least they do if you desire to be consistent.

If you colour model goes as far as to be based on actual physical properties of light, it's extremely complicated to convert RGB to "greyscale". You'll have to dig into spectra and whatnot, for all I know.

If your colour model is centred around CIE L*a*b* as an approximation of human vision, then to be consistent you need to calculate the Y channel of XYZ from your (s)RGB values, and from that next calculate a "greyscale" value. Otherwise, the L* brightness of an RGB value is going to be different from the L* brightness of the grayscale value derived from that same RGB value, and that doesn't make any sense.

If your colour model is totally on the other end of the spectrum, i.e. fiddle with RGB values until you get something pleasing that's fast to calculate, then the .33R+.5G+.17B formula is as good and as bad as any. I've seen R, G, and B being averaged with equal weights, too, and that's no worse nor better. You're perfectly right of course in suggesting that it doesn't make much mathematical sense to calculate a weighted sum of non-linear values, but in this model there is no such thing as 'sense' anyway, mathematical or otherwise, it's merely about doing something that looks to give acceptable results and is fast enough.

Of course, while all colour models are created equal, and all are said to be the textbook standard, we are allowed to have a personal preference. ;-)

Best regards,

Joris Van Damme
info@awaresystems.be
http://www.awaresystems.be/