On samedi 10 mars 2018 21:32:49 CET Matthias Bodenbinder wrote:
> Hi,
> 
> I take the latest postings about denoise issues as an opportunity to ask a
> few questions.
> 
> I am a long term DT user. I have spend hours creating my own denoise
> profiles with several instances of profiled denoise in wavelet and local
> means. But when comparing with rawtherapee or lightroom I have the
> impression that DT is lagging behind. Every once in a while I do a RAW
> development competion with a friend. He is using lightroom, I am using DT.
> When we compare the resulting JPGs, Lightroom frequently beats DT in the
> denoise area. And the same is true for rawtherapee. Denoise is the biggest
> weakness of DT from my point of view.
> 
> It might be a personal issue. May be I just do not know how to do it right.
> But that is my next point. The denoise functionality in DT is just too
> complex from a user point of view. It has denoise bilateral, denoise
> non-local means, profiled denoise, raw denoise and equalizer denoise (do I
> miss something?), And any combination of all of them. This is just too much
> from my point of view. Who actually needs this multitude of denoise
> functions? And who knows exactly what to do with them or how to combine
> them?
> 
> Does DT really need that many denoise functions? I would appreciate just one
> module with a handful of slideers doing the job. Look at Lightroom. By the
> way, is it possible to adapt the denoise code from rawtherapee?
> 
> I would really appreciate if this posting helps to start a serious
> discussion about denoise in DT.
> 
> Thanks
> Matthias
> 
> ____________________________________________________________________________
> darktable user mailing list
> to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org

First, I wouldn't consider the equaliser module as a denoise option (it can be 
used for that, but that's only one use of it, as it uses a very generic 
technique).

*Raw denoise* is applied very early in the pipeline, even before demosaicing, 
so that's one I'd like to see kept in any case.

*Profiled denoise* is the only one NOT considering noise distribution as a 
simple Gaussian, but as the sum of two components, which is especially 
relevant for the darker parts of the image (noise has several contributors, at 
least one of which is *not* related to the light intensity). This module 
requires a camera-specific dataset, so might not be available for all cameras

The other two both assume noise follows a Gaussian distribution (afaik), which 
is (mostly) true for the lighter parts of the image. According to the manual, 
one is especially suited to high ISO images, which might imply the other is 
more useful for medium ISO images (low ISO shouldn't need denoising).

So each module has its specific characteristics and use cases. In fact, I would 
like an _extra_ noise reduction module: "black frame subtraction" (I know 
there's a request in the bug database)...

I'm not sure combining those modules in one module with several sliders is  
possible: they are applied at different stages in processing, and use different 
colour models. And in general, the more of the machinery is hidden behind 
opaque sliders, the less chance the *experienced* user has to get the maximum 
out of the available routines (true, it's easier to get decent results on 
common cases, but styles can handle those). So I'm not really in favor of one 
combined denoise module.

And let's not forget that for any noise reduction method, you'll always have 
to choose between preserving detail and removing noise. I find that high ISO 
images where *all* noise is removed look invariably flat/artificial (esp. skin 
and textiles), whether from DT or any other program. Also, *sharpening* plays 
a big role in how visible the noise is, over-sharpening really accentuates any 
noise present.

Remco


____________________________________________________________________________
darktable user mailing list
to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org

Reply via email to