On Fri, Dec 28, 2012 at 1:17 PM, Christophe Vescovi
<[email protected]> wrote:
> Le 27/12/2012 22:10, johannes hanika a écrit :
>
>> [..]
>>>
>>> A strength of 1.5 (even 2) on chroma is not a problem.
>>> Denoising on (r+g+b), (r-g), (b-g) with a different strength for the
>>> first one could work, and is very simple. I am still not sure if this
>>> transformation have to be done before or after the poisson-gaussian
>>> noise transform (I think before should be ok).
>>
>> your channel r+b+g will have all three variances in it, and r-g would
>> have two of them added up (but the signal part will be much lower in
>> magnitude probably). :/
>>
> I have played a little with the code this afternoon, the results were not
> very different from the r,g,b denoising.
> I have taken into account the variance modification and the wb in the
> transform and have to bias the (r-g),(b-g) channel since the backtransform
> assume a positive value. Well ... it works, but it was still impossible to
> control chroma and luma noise independently, the strength control on the two
> "color" channel having almost no effect, so it is quite useless ;-)

well, i would have been very surprised if it turned out any good..
it's unobvious to my which distribution a sum of (or even difference
between) gaussian/poissonian random numbers follow.

the measured red/blue channels look pretty much like the scaled green
channel (which is how we handle them), but your transform would be
very different. in that context also the unbiased backtransform (which
is different to just algebraically inverting what you did while
transforming to stabilize the variance) doesn't make much sense to me.

i think denoising first and then blending is the safe way of doing
luma/chroma separation.

the other way is to adjust the distance metric when evaluating the
patch differences between points in the nlmeans core. i did a quick
experiments with (L1-L2)^2 instead of <(rgb1-rgb2),(rgb1-rgb2)> but
that didn't seem to work really.

j.

------------------------------------------------------------------------------
Master HTML5, CSS3, ASP.NET, MVC, AJAX, Knockout.js, Web API and
much more. Get web development skills now with LearnDevNow -
350+ hours of step-by-step video tutorials by Microsoft MVPs and experts.
SALE $99.99 this month only -- learn more at:
http://p.sf.net/sfu/learnmore_122812
_______________________________________________
darktable-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/darktable-devel

Reply via email to