Hi,

I've been working on a new noise reduction filter. The working name is "ms nlmeans c noise" and it is in Testing -> Iain Fergusson.

If you could test it and give me some feed back that would be great. You can play around with the advanced options if you want but the top four sliders should be adequate for most images.

Thanks,

Iain

Nice! Here is the original


The fix using default values in your new filter. The small light next to the center deers head is actually a headlight i believe, or a house light.



Really nice results i think. Thanks for the new toy.

Hi Iain
I've just seen your post half an hour ago
I made only 3 tests
so far results are not good, nevertheless I want to post my comments here :
1 - first two experiments made on small very noisy pictures: outcomes not good IMHO
then I tought that I have to try it on a normally noisy photo, because I should not pretend miracles
2 - so, I took a recent photo as generated by the camera, daylight, normal noise; the photo is big even if not dramatically big, 3400*4600 pixels, but after a quarter of an hour I had to shut down Gimp, because it was still working w/o results
I have only 32-bits Win 7, and only 4 GB RAM - thus likely not suitable for this test

I plan to make some additional test using only cropped sections of my recent photos
(cropping should preserve original noise, while downsizing will destroy/altering it)
I'll let you know

Can you please post one or both of these images?
Thanks for testing


Good point. I will see what I can about memory consumption before the final version.

Thanks for testing. Your results so far are useful, even if it means I still have some work to do .

I ran the new filter on this image at 1008x1180 pixels. It took less than 30 seconds on a Win7 desktop (4xAMD processor and 8G RAM). Cropped tight to show detail, and compared to Ian's Noise Reduction filter.

On 3240 x 4320 image, the filter took just under 3 minutes, and the computer's fan kicked up to high speed because the processor was running hot.

Thanks acmespace!

I made some new tests with sections cropped from original image.
I used two diffeernt sizes: 1600x1200 and 3200x2400
Time for running the filter is about 1' for the smaller one, 3 to 5 minutes for the larger one.
I played only a little with the parameters, and -according to my taste- I got good results increasing parameters 3 and 4 from 0 to 2, and parameters 5 and 6 to 3 and 6
With default settings it performs very "delicate" noise reduction.
Anyhow, the problem with my photos is that they are "as produced by the camera" but JPG, not RAW, hence the noise has been somehow "manipulated" by the camera when converting the internal RAW to the external JPG.
About the small very noisy pictures, I will repeat the tests and post the images later.

I decided to post here my very noisy images used in the first test, so that you can yourself see what (and how) they can be improved.



I guess it's useless for now I try to make experiments: you know the filter and you know how to vary the parameters at best

After about 3 seconds with these settings

i got this result. (windows-8.2 64 bit 4Gig RAM).

I made slight adjustments for this before/after pair. This is definitely the best I've ever seen. It took a few minutes on another test with a full image, but not too bad (I'm not hampered by Windows LOL). I have a lot of grainy available-light photos for further testing


5.26
2.0
0
0
Spelling correction, slider 2: I think you mean "coarse" rather than "course"

Nice results there.

Thanks everyone for the tests.

I have made some changes (it's now ms nlmeans c noise2). The automatic noise measurement is now switched off by default. I have also added more sliders so you can fine tune the noise reduction for various tonal ranges of your image. It means you have to fiddle with it a bit more but it works better for more images.

Edit: I also turned of parallel processing for the time being, so it will be a bit slower.

Definitely doesn't like 4x3,000px images. Is there a preferred size?

Thanks for testing the filter.

It's quite memory hungry, so if you have a 32-bit system, or less than 4 GB ram you will have trouble with larger images.

There's a trade-off between memory use and speed. I can reduce memory usage, but it will be slower. I'll think I will make it more memory friendly for the final version. In the meantime you'll will have to test it on cropped parts of you images.

@Sally
you can play with the filter using a cropped size of 3200x2400
performances have improved in systems with 32-bits and 4 gb ram (like mine)
you get the result in less than 2 minutes (it was 3-4 minutes with version 1)

@Iain
made some additional test on a cropped image 3200x2400 (less than 2 minutes)
i tried the default setting (3-1-0-0)
then the Matera setting (5-2-0-0)
then also other two setting: 3-1-1-1 and 4-2-0-0
in my case (with the test image) best result is likely the last one
i post the different versions after having cropped the results not to occupy too much gimp-chat space

..and the last one



as I said, to my eyes this is the best outcome, but everyone can judge
in any case, a very good noise reduction filter, thanks Iain!

not to leave off the small noisy images, I tried again also with those and got good results
the suffixes of the images indicate the parameters used





(originals were already posted)

I'd like to launch a proposal for good photographers, like some around here who from time to time show up:
(assumptions: a very good camera but not so excellent that you never get noise in your photos!)

a- choose a still subject (still life is the most obvious example, but there are others)
b- take many photos of the same subject with different ISO levels: 400, 800, 1600, 3200, 6400 ...
c- download them from the camera in raw format, thus without any camera manipulation
d- you will have different levels of noise: likely none at 400 ISO, just a little at 800 ISO and so on
e- at this point start using the new Iain's filter with different settings
f- take note of those setting that will reduce the noise to one-two stops down, so that (let's say) a photo taken at 3200 ISO has the same noise level of one taken at 800 (or even 400)

by this way you (and all of us, including the author) will have an "objective" estimation of the power of this new filter; so far we had a carrousel of images which are considered OK by the testers -me, Matera, Rod, Acmespace, ...but in any case with a subjective evaluation; the comparicon against an original not noisy photo of the same subject will provide an aseptic reference

something like nowadays is the evaluation of an effective shake-reduction mechanism, which is mainly presented as "being able to gain 1-2-3 stops...".

Iain, I'm interested in what you think about this proposal

Firstly, thanks dinasset for you enthusiasm for my filter.

I like your idea. It will be helpful to see how other people use the filter under different circumstances, and also the results people who don't know the internals can get.

Bear in mind that the filter is not finalized yet. I am going to keep playing with the advanced controls, and probably a small to change how the tonal controls work.