Finer-grained noise profiling, and more

Fer · Post by **Fer** » Mon Oct 27, 2003 3:55 pm

Hello NI staff,

I'm using NI since 2.4, now using 2.6 Pro.
Of course I love it and continuously recommending it on various forums; but having processed some thousands images of different kind (from scanner, from digicams and so on), I noticed a few shortcomings about noise profiling / noise filtering: maybe expressing them here, we may find workarounds, or maybe you may find suggestions for future developments...

First of all, I feel that the rectangular selection is quite a limitation in many cases. Let's present an example.
You did surely hear about new Fuji digicams: S5000 and S7000. Their new SCCD-HR sensors are very noisy, so Fuji implemented a very aggressive in-camera NR, that kills most surface details.
I worked on RAW CCD outputs from those camera with NeatImage, and obtained decent results (please see http://gundam.srd.it/PhotoPages/fuji_dcraw_02.html ); but the RAW CCD output from those sensors (when processed by free tools like DCRAW) is rotated by 45 degrees.
If you straighten it up *before* processing with NI, you'll end up modifying noise characteristics a bit (please see this thread:
http://forums.dpreview.com/forums/read. ... ge=6465075 ), maybe weakening NI action (?).
But on the other side, trying to build a noise profile on a 45-degree test target with a rectangular selection is difficult: you waste a lot of patch space.

Second comment: the noise equalizer has too few frequency bands in my opinion. Some profile have steep variations around low frequencies, i.e., noise is awfully high in deep shadows while quite OK in 1/4 shadows and bearable in midtones. As for now, my equalizer (after fine-tuning the profile with the test target) looks quite coarse, with the 2 lowest sliders to the top (150-200%) and the third one (for example) at 33%.
Now, considering how a frequency-domain filtering works, (you know, not-too-high-order filters and so), isn't a similar situation a bit imprecise?
In particular, I'm observing too much details killing and too few noise reduction in some cases (high ISO shots from newest digicams).
Maybe a finer-grained equalizer would help...

Third comment: High frequency / medium frequency / low frequency noise filtering... isn't it too coarse a subdivision, too?
Maybe more sliders would help figuring a better tradeoff between noise and details?

Fourth comment: sometimes it seems to me, that noise on test target is not modelling real-world pictures accurately enough, expecially for small-sensor digicams.
I build a noise profile (of course I follow your instructions very closely), figure out a noise filtering preset, then apply them to the test target image, and all goes very well & smooth; but when I try the same setup on real-world images (of course same ISO, same device, and so), noise is not cleared "well enough", expecially shadow noise; while quite a lot of surface details (textures) are washed out.
Even trying new values for noise filtering does not seem to help much: seems like noise was somehow too roughly profiled...?
Of course I apply NI on non-processed pics, often straight from RAW conversion (and did build the noise profile the same way).

Last comment: sometimes some kind of "pattern" is visible within CCD noise; expecially with "exotic" sensors like Fuji SCCD-HR. Does NI perform some kind of pattern recognition and suppression, or it's only frequency-based (Fourier-domain band-pass filtering)?

Please note that in the end I'm happy with NI; just noting that maybe with latest digicams I'm pushing it very much... so maybe some improvements could be handy.

Regards,

Fernando

Post by **NITeam** » Mon Oct 27, 2003 6:42 pm

Hello Fernando,

Thank you for your comments and suggestions.

Let me comment on some of them.

1. Yes, the rectangular selection may look like a limitation but this is a feature of our noise reduction method. The rectangular selection cannot be replaced by another shape. On the other hand, I think there is no much degradation (certainly no NI 'weakening') if the image is rotated 45-degrees before noise reduction.

2. Speaking of 'too few frequency bands', I think you are missing the point of the noise equalizer - it has no frequency bands, it has brightness ranges, from shadows to lights. In this context, I do not completely understand your suggestion of a finer-grain equalizer. Probably an example of "too much details killing and too few noise reduction" would better explain it? Otherwise it is mostly a guess-work on my side, I am afraid.

3. I agree that it is perfectly possible to increase the number of frequency range controls. But how many are just right? 6, 9, 20? Considering that people will have to adjust them, the more sliders the more work.

Besides, I have seen no digicam image so far that would really require more gradations. This is my experience.

4. Did you try to build profiles from an image from the same series as your working image? Maybe the profile you used simply was inaccurate. Are you sure the device mode of test image really matched the device mode of the working image?

Also, I would suggest you to always use your own judgement in adjusting filter settings instead of relying on the ready-made presets. It is my impression that you most use the ready-made presets and adjust the filter yourself only in rare cases.

The last point: Neat Image is not frequency-based (Fourier, etc.) and can peform some kind of pattern recognition sufficient to deal with most CCD noise. However, I would not be surprised by anything related to Fuji sensors because to me it often seems that Fuji engineers outwitted themselves with those 45-degrees things. Nevertheless, I am not aware of any trouble with processing those images in Neat Image.

Thank you for your comments and suggestions. All in all, I summarize them as give us more control and make the noise models more accurate. I am not sure about the utility of even more control than is available now, but we undoubtedly like the idea of improving the accuracy of noise models used in Neat Image and will further work to this end.

Best regards,
Vlad

Fer · Post by **Fer** » Mon Oct 27, 2003 7:28 pm

Hello Vlad,

let me clarify a bit..

1) Ok, I'll use rectangular selection then!

2) From what I understand from the manual, leftmost sliders affect "darker" colours, so, for example, shadows: where the noise is most disturbing (expecially chroma noise). Moreover, building a profile with the test target often results (just tried with a Powershot S30 @ 400 iso) in the leftmost two sliders at the maximum (sometimes they even reach 250%, that is a value not selectable by hand), while the third-leftmost is positioned quite lower (40-50%). So a steep adjustment indeed... having a finer-grained "shadow" sliders, we may find, say, that settings like (from left to right) 250% 180% 90% 50% could be achieved, with smoother control over the noise/details tradeoff... you know, with digicams, bright colors seldom show noise, so sliders are even too much for the brighter end of the gamma: noise tends to show up mainly in underexposed areas (it's the nature of the beast).
3) I think a couple more sliders would do...
4) Yes, absolutely: I used identical settings for both the test target shot and the real-world pics. And, it's not easy to build a profile on real-world images for a 3mp camera: you won't find uniform areas large enough to fine-tune the profile other than for high-freq noise.
And, I don't use presets at all (I always find they sacrifice too much details)...

Last: Fuji SCCD-HR noise is peculiar indeed: I've seen many RAW images (so no in-camera NR) and they show an *incredible* amount of noise (even at the lowest ISO!) in shadows and underexposed areas. This is why Fuji adopts such an heavy in-camera NR, that destroys lots of surface details (forums are full of complains about that). Maybe, if you have the needed time, you could have a look at some of those images? I failed to filter them efficiently without destroying too much details, but I feel that with some minor tweaks to some internals (or simply with a good advice from you) I could find a viable setting (and of course contributing it to the community).
Those new Fuji camera are very good indeed, features and price-wise; but too much noise and too aggressive NR are causing people to stir away from them (and online reviews to stigmatize them).

Thanks for your attention, regards!

Fernando

Post by **NITeam** » Mon Oct 27, 2003 10:02 pm

Thank you for the clarifications and additional comments.

2) Yes, now I understand. Earlier it was not clear because 'frequency' and 'brightness' were intermixed.

Regarding the equalizer I would not worry about the number of sliders because inside we anyway use somewhat more accurate representation.

And an advise: if you end up with 250% in some sliders, try to start with another (darker in this case) initial analysis area to avoid going out of ranges.

3) Ok, we will take this as a suggestion to consider. You are the first to ask.

4) Then please make sure (1) the profile is good (without those 250%, etc.) and if still unsuccessful (2) build a profile on the spot from the working image (or a special shot of the calibration target printout). If the quality is still not improved then there is probably something unusual about the camera. I would want to see an example.

Last: Thank you for the suggestion to look at the latest Fuji images. I think our team will take a look at these, just have to find good sources of ready-made RAW Fuji images and tools to work with them (unfortunately (or otherwise), the Fujifilm cameras available to us are not based on SCCD-HRs). So, if you have good links to such websources, I would appreciate if you shared these with us, either in this forum or by e-mail.

Thank you once again,
Vlad