Rod, I know the extreme compact style of David, for him it's certainly not a matter of massive coding (only of feasibility in GMIC).
Unfortunately, he did not yet enter this discussion, likely busy.
Another G'MIC noise expert is Iain, if he reads from time to time Gimp-chat posts maybe he can enter the discussion.

Dinasset, do you want a fast filter (almost real-time) or a filter that takes a quarter of an hour?

Rod, altough I'm not a programmer, I think this is really difficult and a lot of coding. Estimating a noise curve from a single image isn't simple. We'll bet? After estimating the curve, the curve has to control (steer?) a filter (which one?), also complex. So I'm curious too.

what I would like to have from a performance viewpoint?
more or less: one minute delay per megapixel of input image
id est:
- 1 minute for an image 1000x1000
- 4 minutes for an image 2000x2000
- 9 minutes for an image 3000x3000
Is that reasonable?
(with an average PC, obviously, not with a 24mbRAM high performer monstre; for instance a PC with Intel Core I5 and 8mbRAM

Dinasset,

In G'MIC, there is a command:



So, this command is for some kind of noise estimation, I suppose. It is used in the following filters:

Repair>Smooth [iuwt]
Repair>Smooth [block pca]
Repair>Smooth [wiener] (David, Jérôme, does this filter work correctly?)

Of course you have to adjust the values. But with every noise reduction sofware you have to do that, to your personal taste. (Also if they promote it is fully automatic)

In my theoretical view, some G'MIC expert should then able to write a filter which:
1) runs that command
2) according to the results, calls the most appropriate G'MIC Smooth filter (one of the existing ones or a "compilation")
If this is not possible, for me this means that the estimation does not really "understand" which denoising method is to be preferred.
In G'MIC we have so many...

Dinasset,

Maybe you can test the "Smooth [block pca]" filter with a lot of images. This filter is fast enough to give a chance (faster then patch-pca), I think. It is very advanced.

Perhaps, with the correct setting, it does what you want.
I would love to have a tutorial for that filter.

iarga, can you provide a set of noisy images?
my collection has not many, because I do not use normally high ISO.

Dinasset, I can't, because I don't have really good images to test with. I'm so sorry. Maybe a search on internet can provide something?

Also you can test with adding noise yourself, with noise generators.

I could probably take a few, my Pentax K10D is old and quite happy to add noise

Kevin

done some tests, "Smooth [block pca]" behaves badly IMO.
I guess it's a loss of time, there are better G'MIC smoothing filters IMO (for instance the most famous: anisotropic, which is also very quick)

FWIW I tried playing with the "Smooth [block pca]" filter when it first came out. I couldn't see any difference using the filter. I had started playing with Linux at the time and decided to try Gimp in my Linux VM. Here the same filter worked and did an great job. I have never been able to get this filter working with Win7. Probably something with my system, but it is something to keep in mind.

bilbo, did you set this strange behaviour to the attention of David?
G'mic should behave OS independently

Dinasset, The preview isn't comparable with the result. So don't watch the preview. I did some testing. I think this is a nice filter. But it gives wrong colors. Maybe with YCbCr [luminance] or Lab [lightness] it would be better?

I will try.

I also think that this is (much ) too slow for your purposes.

I didn't talk based on preview, I tested it.
Look at these 3 examples of outcomes (source is a noisy image on Google)

block-pca with default parms


block-pca with parameters values almost doubled


anisotropic with default values


up to you to judge (IMO anisotropic, even w/o tuning, is much better, but that's my opinion)

Dinasset, with such big images I won't dare to try block-pca . The default anisotropic filtering is nice, maybe a bit smooth, but I understand what you want to show. In this image without tuning, it is okay, but with more noise, you have to tune. With eye, hand, mouse, monitor and human brain, you then can do a nice job.

I did some testing:
With only YCbCr [luminance], "block-pca" does a better job. But I now see different patches. I don't know how to have a better overlap.

After some testing, I prefer Smooth [uiwt], for general purpose (from the "noise_std" filters).
I tested this filter with extreme noise, a scan from a printed image. In the code I changed the maximum treshold. After that, with this kind of images, the result is very good.

original:


block pca :


uiwt:


gaussian blur, radius 16:


anisotropic:


The noise estimation in this filters is different from what you mean. Your question isn't answered. "Is full automatic noise reduction possible in G'MIC?"
And I think another question: "Is there a possibility to automatic control the "Smooth [anisotropic] filter" or other filters in G'MIC with some kind of noise estimation?

yes, iuwt is nice
your last question is exactly my open unanswered question
no G'mic people entered this topic so far

I think uiwt with that treshold is about the same as a gaussian blur( see above) So I think uiwt has to work with a lower treshold.

Of course this isn't a good test-image at all. You better take Barbara or Lena.

I changed the title to ask David whether he has an opinion on the possibility for G'MIC (which has a lot of excellent denoising filters) to create a sort of a front-end to "understand" the noise present in the input image then call the (approx) best G'MIC solution (filter and parameters).
Thanks

I would have a look on Neat Image.

They seem to employ a simple denoising algorithm yet have great results due to their Noise Profiling.

I guess Noise Profiling is all about locating good patches with no edges in them, calculate their Mean Value and their STD.
If the algorithm found many different patches with different Mean values a Noise Curve can be generated.

If one can create "Optimal" parameters of the denoising algorithm given the Noise STD than this curve can make the output significantly better.

Actually this means the algorithm spatially adaptive to the pixels.

Yes, there are probably ways to do this in G'MIC. I'm sure of that.
Now, as always, is there a contributor with enough time to tackle this issue ? Definitely not sure