[argyllcms] Re: noise and gamut, input profiles with B2A table

Thanks Graeme, Nikolay and Alastair for sharing your thought and
experiences.

Nikolay Pokhilchenko wrote:
About great gamut difference with denoising: I do downsampling for gamut computing more often, then denoising, because most of denoise filters are trend to desaturate noisy areas, especially in darks. Denoising is prescribed in case of sparkling noise, for example at long exposures. 


Klaus, may be Your image is very noisy, as I've seen in
normalVsDenoised.wrl?


No, it's not /too/ noisy, but it's very textured and the gamut surface
is very jagged. It's a Anagramm scan of an acrylic painting on canvas: a
women in a blue dress. The dominant element is the blue dress. It
contains /very/ saturated shades of blue: there are both very dark and
very luminous shades. Some of them tend towards cyan/green, others
(especially the darker ones) are /slightly/ purplish. The furniture in
the background and the (already very pale and bright/yellowish) skin
tones seem to withdraw from the dominant blue, but when you take a
closer look, you will notice that several strokes contain also some
thoroughly saturated colors (red, yellow, green). I don't know exactly
what already happened with the data, even though I know much about them:
I saw the original, I saw teary softproofs (every rendering intent of
the basICColor output profile yields a different disaster), I saw some
testprints (Epson 11880 on canvas) -- and I tried to console and advise
the desperate operator: this painting is a real challenge. He gave me
the Photoshop file (only a representative detail, the whole psb is >
2GB). I guess the bottommost layer contains nothing more than the
carefully sharpened raw scan. Above is a linearization layer (generated
by dint of a reference gray wedge) ... and a /bunch/ of adjustment layers.

I decided to turn off all adjustment layers except the linearisazion and
to start from scratch with argyll -g profiles)

It's hard to judge the results on screen without seeing original, the
more so as many of the source and destination colors are out of gamut --
but inasmuch as I can assess it, even the with the very first argyll -g
results dwarf anything else, including the result the skilled operator
obtained after /hours/ of correcting.


Alastair M. Robinson wrote:

You have me wondering now about thresholding as a means of ignoring
outliers due to noise. Something to ponder. :)


Graeme Gill wrote:

It's a reasonable thought if typical images have such outliers. There
are many possibilities though. For instance an almost equivalent idea would be to alter the gamut mapping so that a percentage of the source gamut ends up getting clipped. This might be possible currently by creating a new gamut mapping template in xicc.c. 

But how to determine the right percentage? I think that clipping a fixed
proportion of the gamut can be an option for "average" situations, but
it's less advisable in my case: Denoising and downsampling affects the
less important background colors and highlights of the brush and canvas
textures much stronger than the dominant blue-in-blue area, and that's
the right thing: Clipping colors only apeearing in parts with high
spatial frequencies apparently improves the result.

Nikolay Pokhilchenko wrote:
Usually I do downsampling to 12.5-25%. I'm determining the number of downsampling by smallest colored details, which I want to preserve good mapped. But I'm converting the original image without denoising. 



The differences between profiles with downsampled and denoised gamuts
seem to be subtitle on screen, even though the vrml files suggest larger
differences. I think I have to make test prints to judge the results.

Graeme Gill wrote:
There are many different filters that you might apply to the image. Downsampling is one, which groups colors that a spatially close and averages them. Another would be a quantizing algorithm, that groups colors that are close together in color space and averages them. (I'd be inclined to look at the latter first - I started work at one stage on a program to create a test chart based on an images colors using median cut quantization algorithm, borrowing it from the one I created for xli, and the same approach could be used). 

Great inspirations --- thanks, Graeme! :-)

Using a median cut quantization algorithm for gamut mapping purposes
sounds promising. I also love to hear that you are experimenting with
image specific test charts. Did you also think about doing this in
spectral space? -- That's a topics I'm look forward to deal with in the
(hopefully near) future.

Klaus

Other related posts: