Project

General

Profile

Bug #9883

Chroma noise reduction does not take exposure correction into account

Added by Istvan Kovacs about 4 years ago. Updated about 3 years ago.

Status:
New
Priority:
Medium
Assignee:
-
Category:
Darkroom
Target version:
-
Start date:
03/31/2014
Due date:
% Done:

0%

Affected Version:
git development version
System:
unknown
bitness:
64-bit
hardware architecture:
amd64/x86

Description

Enabling wavelet-based profiled noise reduction with blend: uniformly, blend mode: color on high-ISO images causes ugly noise to appear when high exposure correction is used. Luma filter (non-local means + blend mode: lightness) does not show the problem.
Below is an extreme example.
Note: this was an unusable image. It was about 3 stops underexposed, shot at ISO 3200. Now, applying an exposure correction of +3 EV, DT still tries to use a noise profile for ISO 3200. Applying wavelet/colour noise reduction with the ISO 3200 image creates a lot of noise (top half of top image). Manually setting the noise profile to ISO 25600 (ISO 3200 * 8, or +3 EV), the noise is reduced without creating luma artefacts.

History

#1 Updated by Ulrich Pegelow about 4 years ago

  • Tracker changed from Bug to Feature

Moved to features.

Not a bug really. What you seem to want is a mechanism within noise reduction that takes EV compensation into account.

#2 Updated by Istvan Kovacs about 4 years ago

Well, if turning on noise filtering adds noise to the image, I'd say it's broken. I mostly use NoiseNinja for denoising, so it's not a big deal, as far as I'm concerned.

#3 Updated by Ulrich Pegelow about 4 years ago

  • System set to unknown
  • Affected Version set to git development version
  • bitness set to 64-bit
  • Category set to Darkroom
  • Tracker changed from Feature to Bug

Seems I misunderstood. Introducing new noise as part of noise reduction is likely a bug. Could you please supply your example image plus XMP?

#4 Updated by Istvan Kovacs about 4 years ago

Please check http://tech.kovacs-telekes.org/files/darktable-issue-9883/
Note the green dots that appear at 100% zoom; set profile manually to ISO 25600, and they'll be gone. To make it even more evident, you can increase exposure correction beyond 3 EV.

Thanks for the investigation.

#5 Updated by Igor Kuzmin over 3 years ago

I can confirm this with Canon 7D images, 1600 ISO is noisy enough for it to manifest. Hot pixels module helps reduce this issue, but some dots are still there. It's not very critical for me since downscaling smooths them away. Increasing the strength or choosing preset for larger ISO removes the dots too, but at the cost of some details in picture. I don't think that exposure correction has much to do with this, the problem is probably just an artifact of algorithms - wavelet profiled denoise with color blending mode makes luma noise more pronounced (actually I would say it increases sharpness overall, even show more details, but these dots are annoying nonetheless).

#6 Updated by Pascal de Bruijn about 3 years ago

(sidenote)

@"What you seem to want is a mechanism within noise reduction that takes EV compensation into account."

This is probably a very very bad idea... All our modules have clearly reproducible behaviors (and I think it should really stay that way). Making module behaviors dynamic would likely become a great source of confusion.

If anything, if there are weird/extreme interactions between modules, reordering them might be considered.

#7 Updated by Istvan Kovacs about 3 years ago

Pascal,

You're right; I fell into the error of proposing a solution to an issue I do not fully understand. Please feel free to edit the bug description (if that is possible). My issue is that applying chroma noise reduction increased noise, that's the only thing that I, as a user, know for certain.

Thanks,
Kofa

Also available in: Atom PDF