Nikon 2D FFTs Showing Involuntary Noise Reduction

[bclaff]

This isn't really a new topic but I have some data on more recent camera models and I find the visual differences between the models "interesting".

Background: despite turning off all Noise Reduction settings it appears that all Nikon DLSRs do some sort of signal processing that amounts to Noise Reduction (NR) when model dependent exposure time is reached.
A 2D FFT can give us some (vague) idea as to the symmetry and strength of the processing.
(For the actual strength, examining the measured noise reduction is far more objective!)

Each row shows four 2D FFT power spectrums displayed with a logarithmic scale (from ImageJ).
Each image is from a 256x256 center crop of the Gr channel from a raw image.
Each image is a "lens cap" shot take at the lowest numbered ISO for the camera with all NR settings off.
Left leftmost image is at the camera's fastest shutter speed.
The second image is for the next shutter speed that does not exhibit additional signal processing.
The third image is for the first shutter speed that exhibits additional signal processing.
The last image is at the camera slowest shutter speed.
Each row is labeled with the camera model and exposure threshold.

D70 (1s)
D200 (1s)
D300 (1/4s)
D700 (1/4s)
D90 (1s)
D7000 (1/4s)

If you're not experiences at "reading" FFTs, I suggest a quick look at Visually Interpret FFTs

I invite your comments and withhold my interpretation, for now.


Bill

[bjanes]

Each row shows four 2D FFT power spectrums displayed with a logarithmic scale (from ImageJ).
Each image is from a 256x256 center crop of the Gr channel from a raw image.
Each image is a "lens cap" shot take at the lowest numbered ISO for the camera with all NR settings off.
Left leftmost image is at the camera's fastest shutter speed.
The second image is for the next shutter speed that does not exhibit additional signal processing.
The third image is for the first shutter speed that exhibits additional signal processing.
The last image is at the camera slowest shutter speed.
Each row is labeled with the camera model and exposure threshold.

Bill,

The findings are indeed interesting. All of the sensors exhibit attenuation of the high frequencies to a varying degree. The effect is most noticeable with the D300, where it is greater at 1/4 sec than the longest exposure.  Similar findings are present with the D700 and a line is present horizontally, indicating pattern noise (possibly due to parallel readouts). With the D90, the filtering is most prominent in the longest exposure. The D7000 (the newest sensor) exhibits less filtering and I don't see much difference between 1/4 sec and the longest exposure. It also shows slight pattern noise.

Interestingly, the patterns were difficult to ascertain on my laptop which presumably has a TN screen with rather marked directional variance in luminance, but were better seen on my desktop using IPS technology.

Does the D300 where the effect is most prominent use CCD or CMOS technology? The D70 and D200 do use CCD and exhibit some pattern noise and relatively little filtering, similar to the D7000 which uses a later generation CMOS.

Regards,

Bill

[ejmartin]

Marianne Oelund has done an extensive analysis, see

http://forums.dpreview.com/forums/read.asp?forum=1021&message=34309201

[bclaff]

Emil,

Yes, I've read that thread more than once and found it quite interesting.

Marianne, devised, constructed, and exposed a very specific target to investigate the algorithm used.
It's a very interesting investigation.
I think a detailed comparison of her work with the patent US 8,009,207 B2 Noise reduction device for reducing noise in image using blackout image, electronic camera, program, and method would be called for.

My investigation is related but different.
Lens cap shots don't reach the threshhold for the algorithm investigated by Marianne; yet something still is happening.

Regards,
Bill

[bclaff]

I'm a little surprised that no one has picked up on the dots along the x-axis of the D700 images.

I'm attaching the 1/8000s crop as a hint.


Bill

[Bart_van_der_Wolf]

This isn't really a new topic but I have some data on more recent camera models and I find the visual differences between the models "interesting".

Background: despite turning off all Noise Reduction settings it appears that all Nikon DLSRs do some sort of signal processing that amounts to Noise Reduction (NR) when model dependent exposure time is reached.

Hi Bill,

Thanks for sharing your findings. I'm not sure if it possible to test this with only the body cap instead of a lens. There are cases where the camera and lens electronics caused some sort of electronic interference, which could cloud the other effects.

Personally I would prefer to also test this behavior on actual exposures, because we have no certainty that the noise floor and the actual image data are treated the same way, they probably are but we can only assume that until it's actually tested. This will probably require to use a lens for some camera models, but given the higher Photon shot noise levels it will be less of an issue. I've also visually noticed that shadows and highlights can undergo different levels of in camera noise reduction, so the tests should also include a few exposure levels (e.g. as metered of a uniform surface or through a sheet of opal glass or similar acrylic, and 2 stops over and under, to avoid clipping the noise spectrum). One could also shoot a white noise target, but then the optical (MTF) influence is added to the equation.

As an alternative presentation you could try the ImageJ plugin called "Radial profile" (angle), which produces a graph of the power at various radii on the FTT frequency or power plots.

Cheers,
Bart

[bclaff]

Bart,

I collect data of all sorts from willing collaborators who own camera models that I do not.
In this analysis, some are lens cap shots and some are body cap shots.
The D700 are body cap shots contributed by M. R. Blume.

I agree that signal level affects the signal processing.
And that studying the effect at "normal" signal levels might be more photographically relevant.
Others, like Marianne Oelund, have done work in that area.

Unfortunately, the best way for me to collect consistent data across multiple contributors is to stick with "cap" shots.
(Although you're given me an idea about the reuse of images collected for gain analysis.)
I do think this gives us some information about the low signal characteristics of the cameras since the cap shot is the "limit" case for low signal.

Thanks for the ImageJ tip (I'm not a power user).
I'll try that out.

Regards,
Bill

[RFPhotography]

What does all of this mean for real world shooting?  The analysis of the involuntary noise reduction is, perhaps, interesting but if it can't be turned off then how does it impact shooting in a practical sense?

[bjanes]

What does all of this mean for real world shooting?  The analysis of the involuntary noise reduction is, perhaps, interesting but if it can't be turned off then how does it impact shooting in a practical sense?

It does make a considerable difference in astro-photography where the noise reduction can erase a dim star from the field. See Christian Buil. It also helps to satisfy scientific curiosity and answer a frequently asked question: are raw files really raw? For practical terrestrial photography, it probably makes no difference and you can ignore the thread if you wish, but why denigrate the efforts of those who are interested in the topic?

Regards,

Bill

[RFPhotography]

Your personal disdain for me continues to show, Janes.  There was nothing in my comment that denigrated anyone's efforts.  I simply asked a question.  Two questions, actually.  I understand the scientific curiosity point.  As far as RAW files being RAW (or not), that's been known for some time.  The astro-photography issue is interesting and one I hadn't immediately thought of - so the point of the question was valid.  And yes, despite your assertions to the contrary, I'd consider that a practical shooting matter. 

[bclaff]

Bob,

I don't feel denigrated, and if I did I'd take it up with the poster offline.

Although the effects are "known" the effects differ by camera model so continued testing is called for by those who care because you cannot assume that a new model does "it" the old way.

I think the novel aspect of my post is that you don't generally see this information for multiple camera models side-by-side.


Bill

[bclaff]

The dots along the D700 horizontal line are the result of vertical banding.
The number and placement indicate that something repeats every 6 columns.
(Since we're only looking at one channel, that's every 6 pairs of columns.)
The banding arises because the output of the parallel column amplifiers is not even enough.

The D700 owner sees nothing unusual in his photographs.
Probably the banding is only visible/detectable in very dark regions (for his camera).

Although the banding seems evident in the image posted in reply#4 I find the 2D FFT, which nails the frequency, more "objective".

:-)
Bill

[madmanchan]

Many camera models (not just Nikon) do internal dark frame subtraction or other field calibration once a certain exposure time is reached.

[bclaff]

Eric,

Yes, but except for the D7000, the FFTs indicate something other than (or in addition to) frame subtraction is going on.
For the D7000 I'm guessing a form of frame subtraction and mild filtering.

Regards,
Bill

[joofa]

It does make a considerable difference in astro-photography where the noise reduction can erase a dim star from the field. See Christian Buil.

But, then, IMHO, the solution is to use the right camera, which is designed for such tasks, and there are scientific-grade cameras out there that just do that. The quoted author should do a more thorough investigation rather than a few Canon and/or Nikon brands.

Sincerely,
Joofa

[joofa]

Although the banding seems evident in the image posted in reply#4 I find the 2D FFT, which nails the frequency, more "objective".

Yes, FFT can be helpful here sometimes. I downloaded the following image from the Internet, which exhibits some banding,  and darkened the vertical and horizontal axes of the FFT plot using the brush tool in Photoshop. After IFFT the banding seems to disappear:

For the brave who want to do such FFT analysis in Photoshop (CS3) on a Mac, I have a freely available FFT+IFFT filter set available on the website in my signature, which will give you the exact same dimensions of FFT plot as the spatial image, unlike ImageJ that gives a plot to the next highest nearest power of two. And one can also use the marquee selection tool to do a localized FFT analysis as the image below illustrates with:

Joofa