RE: [APML] Synthetic Luminance Channel for Film Images

[Thor Olson <Thor.Olson@EFI.COM>]

> -----Original Message-----
> From: Jerry Lodriguss [mailto:jml@astropix.com]
...
> In my original post with the subject line "ESO Tarantula CCD 
> Image" I put 
> forth the opinion that creating a synthetic luminance image from the 
> individual RGB images would not gain any increase in signal 
> to noise ratio 
> when composited into an LRGB image.
> 
> I did some experiments that seem to prove my opinion wrong.


Hi Jerry,
I like your first opinion better, but it kinda depends on how you
define signal to noise.  I downloaded your test image after reading
your web page experimental results and tried to understand why you
would see the differences you described.  I have some ideas that
might explain some of your results.

I looked at the 50x50 corner from your image and converted it
to grayscale.  Though my statistics matched yours for the color
image, they differed in the grayscale result.  Depending
on what I selected as my RGB color setup, I could get different
answers for mean and standard deviation of the grayscale converted
result.

I think what you may be seeing is the effect of using different methods
to convert rgb to luminance.  The conversion is just a weighted average
of the three channels, so there are as many ways as there are weighting
combinations.  Some common ones are equal weights (not a good match
to human response, but mathematically simple) which your
33% weighting represents.  Another common one is the video convention
of 30% red, 59% green and 11% blue.  This is closer to a visual
brightness weighting.

If you look at the statistics of your three channels they have:
       mean    std
red    19.88  4.77
green  30.87  5.67
blue   42.53 11.24

I don't think it's quite proper to divide the mean by the standard deviation
for an object that is supposedly "black", but that's a different topic.
Just looking at the std numbers, blue is obviously the noisiest channel,
so any luminance formula that places more weight on blue will tend
to give a noisier result.  In this case equal weights gives std=5.05,
the video weighting gives 4.12.


 
> Can anyone find a better way to average the three individual 
> color channels 
> to create a synthetic luminance image with even lower signal to noise?
> 

Try the weights 53%, 37.5%, 9.5% on the red, green, and blue channels.
You will get a standard deviation of 3.7.  It's probably no further
from being a "visually correct" luminance than the equal 33% weightings,
but it results in the lowest possible noise for this image.


I don't happen to believe that "luminance layering" improves
the signal to noise of the result without distorting the orginal
color information.  It essentially moves the noise around, taking
advantage of the human visual systems relative blindness to high
frequencies in the chroma channels.

But since truly calibrated colorimetric images are so difficult,
we've already crossed the "correct color" boundary and should be looking
for ways to make attractive (but uncalibrated) pictures.  And this
seems like another tool to apply to the cause.  If you want to 
know the weights to optimize any 3-channel image let me know, I can
describe the recipe offline.

Thor.

Thor Olson                            
web-res astrophotos at:                    www.nightscapes.net

   "Man has ten, computers have two, Nature has e fingers."






--  APML Archives at <http://astro.umsystem.edu/apml/>  ---
             Unsubscribe at <majordomo@seds.org>