Project

General

Profile

Bug #8345

White outlier noise pixels

Added by brteag00 - about 8 years ago. Updated about 8 years ago.

Status:
Fixed
Priority:
High
Assignee:
-
Category:
General
Target version:
Start date:
Due date:
% Done:

100%

Affected Version:
System:
bitness:
64-bit
hardware architecture:
amd64/x86

Description

Darktable MASTER is producing scattered white pixels. I originally thought it was the exposure module, because the problem gets worse (ie, more white pixels) when the black point is increased. See white_pixels_1.png (full crop), attached. The problem seems to be particularly pronounced in images with a lot of red.

However, after disabling all the modules, there are still white pixels. See white_pixels_2.png (full crop). This suggests the problem might be in the rawspeed import code, somewhere. Interestingly, as the black point is increased, the pixels that are rendered white change.

An example RAW is attached.

Camera: Canon 50D
Codebase: git MASTER

History

#1 Updated by brteag00 - about 8 years ago

Because of the 256Kb file size limit, I can't attach the files to the trac bugreport. They're in my dropbox Public folder.

Example RAW file:
http://dl.dropbox.com/u/387385/IMG_2371.CR2

Full crop 1:
http://dl.dropbox.com/u/387385/white_pixels_1.png

Full crop 2:
http://dl.dropbox.com/u/387385/white_pixels_2.png

#2 Updated by Igor Kuzmin about 8 years ago

I have 400D and have the exact same problem. So this bug isn't limited just to 50D.

#3 Updated by Anonymous about 8 years ago

I can confirm this too. (EOS 50D, ISO400 and night shot with long exposure)

#4 Updated by Simon Spannagel about 8 years ago

I have the same problem (Canon EOS 550D), but some pixels show up as purple or blue instead of white. This happens not with all but with a lot of images, especially in dark regions of the picture and is - as described - amplified by using different modules.
In my case you really can see it while setting black point, rotating pictures or using the eq. II for any action.

Example File: http://rapidshare.com/files/439312828/rotation-artefacts.tar

#5 Updated by Igor Kuzmin about 8 years ago

400D also produces red noise when shooting in dark places with high ISO settings, but it's different from this white noise. Median filter in demosaic plugin gets rid of red pixels completely, while though it reduces amount of white pixels, they still remain even when you turn the edge threshold up to 1.0 (ignore edges). So IMHO red noise is really there, and white pixels are some artifacts of image processing.

#6 Updated by Johannes Hanika about 8 years ago

i just pushed a new median filter. same place: demosaic edge threshold slider. try values of 0.001 or so.

also note it's only on for performance reasons (the cpu version is really slow!) if zoom > 0.5 (i.e. when demosaicing is actually done) or during export.

this removes all outlier noise for my low-light 400D images, please test!

how it works: it's an edge aware median filter, which splits up pixels in an 3x3 neighborhood in two populations. only the one the center pixel belongs to is used to calculate the median.
new exception to that: if the center pixel population consists of only one member, the regular median is used for this pixel.

it's unrelated to rawspeed (libraw version suffers the same, i tested), so i strip this from the summary :)

#7 Updated by Simon Spannagel about 8 years ago

Still the same issue for me, try it using the same attached file as above.

#8 Updated by Johannes Hanika about 8 years ago

Replying to [comment:8 simn]:

Still the same issue for me, try it using the same attached file as above.

actually i forgot to push. commit 35aa8bb5073a5b6c63f884d7f919da4ba5d7bb10 should be the one now.

#9 Updated by brteag00 - about 8 years ago

  • Status changed from New to Fixed

Works for me, even with the edge threshold set to 0.0. (Sorry to have mis-identified it as a rawspeed issue, I forgot that the demosaicing module was still in the pipeline.) Thanks for the fix, Johannes!

Closing the bug. @simn - if this doesn't resolve what you're seeing, you might open another bug.

#10 Updated by Igor Kuzmin about 8 years ago

Replying to [comment:10 brteag00]:

Works for me, even with the edge threshold set to 0.0. (Sorry to have mis-identified it as a rawspeed issue, I forgot that the demosaicing module was still in the pipeline.) Thanks for the fix, Johannes!

Closing the bug. @simn - if this doesn't resolve what you're seeing, you might open another bug.

White pixels are gone, yes, thanx for that.
But 0.0 setting supposed to disable median filter. Now there are two 0.0 edge threshold settings - one with median filter enabled, another with disabled. For example set edge threshold to 0.001, then press "<" (left of 0.001), edge threshold is displayed as zero, but median filter is applied. Press "<" again, and filter gets disabled, all noisy pixels are here again. Press ">", and edge threshold becomes 0.001, filter is of course enabled. It's quite confusing. We probably need separate checkbox in demosaic plugin for enabling/disabling median filter.

#11 Updated by Johannes Hanika about 8 years ago

Replying to [comment:11 parafin]:

But 0.0 setting supposed to disable median filter. Now there are two 0.0 edge threshold settings

gah! no more controls please!

this is just a rounding issue.. you can't subtract 0.001f and expect it to be 0.0f, so now these small values are clamped to 0.0f. makes the behaviour somewhat more consistent indeed.

Also available in: Atom PDF