D800 and D4 Pattern Noise Analysis with FFTs

Started Apr 2, 2012 | Discussions
bclaff
Contributing MemberPosts: 672
Like?
D800 and D4 Pattern Noise Analysis with FFTs
Apr 2, 2012

Certain types of pattern noise and signal processing can be detected using 2D FFTs (Fast Fourier Transforms).

I performed such an analysis of dark frames (raw) from a D800 and a D4 using ImageJ.
Here's the D800

And the D4

The ISOs from left to right are ISO 50, ISO 1600, ISO 6400, ISO 25600, and ISO 102400 (D4).
The images have been darkened to accentuate the patterns.

Sometimes I find it helpful to view the images from several feet to see subtle features.
The "ideal" FFT would be a uniform gray.
A good guide to interpreting FFTs is here:
http://qsimaging.com/ccd_noise_interpret_ffts.html

Strong vertical or horizontal lines general indicate pattern noise in the horizontal or vertical direction respectively. Both cameras exhibit some pattern noise.

Little "ticks" along an axis is usually a sign of banding, often due to unbalanced channels in a multi-channel readout scheme. This seems very well controlled in both cameras.

The classic "spherical" look of signal processing is absent. There is no evidence of additional signal processing being applied to the raw. In Nikon cameras that general takes place at longer exposures such as 1/4s and longer and will be the subject of a separate study.
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Nikon D4 Nikon D800
If you believe there are incorrect tags, please send us this post using our feedback form.
Macgupta
Regular MemberPosts: 134
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to bclaff, Apr 2, 2012

In the ISO 102400 I see a sort of square with rounded corners, quite circlish. Could be an optical illusion. On the other hand, I just calibrated my monitor.

Reply   Reply with quote   Complain
bclaff
Contributing MemberPosts: 672
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to Macgupta, Apr 2, 2012

In the ISO 102400 I see a sort of square with rounded corners, quite circle-like.

You know, I think you're right. Much more subtle than the usual "sphere".

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Reply   Reply with quote   Complain
semorg
Contributing MemberPosts: 938
Like?
Bookmarked (nt)
In reply to bclaff, Apr 2, 2012
-- hide signature --

search engine for photographers:
http://www.f8daily.com/search

Reply   Reply with quote   Complain
Dave Largent
Senior MemberPosts: 2,068
Like?
I see it faintly
In reply to bclaff, Apr 2, 2012

bclaff wrote:

In the ISO 102400 I see a sort of square with rounded corners, quite circle-like.

You know, I think you're right. Much more subtle than the usual "sphere".

I think I see it, too. What does it mean?

Reply   Reply with quote   Complain
Flashlight
Veteran MemberPosts: 5,216
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to bclaff, Apr 2, 2012

I think I see this (very faintly):

bclaff wrote:

-- hide signature --

Philip

Reply   Reply with quote   Complain
Dave Largent
Senior MemberPosts: 2,068
Like?
oh yeaaahhhhh
In reply to Flashlight, Apr 2, 2012

Yeah, I think I see it. Unless my eyes are playing tricks on me based on what you said.

Reply   Reply with quote   Complain
Cipher
Senior MemberPosts: 2,275Gear list
Like?
Aren't they reversed?
In reply to bclaff, Apr 2, 2012

The top one's dots seem coarser so it looks like from a lower resolution camera? Like the D4 and the bottom one looks finer so it should be the D800? I'm probably wrong...but maybe you can shed some light on it.

 Cipher's gear list:Cipher's gear list
Sony RX100 Nikon D700 Panasonic Lumix DMC-GF1 Olympus OM-D E-M5 Nikon D800 +13 more
Reply   Reply with quote   Complain
Iliah Borg
Forum ProPosts: 15,670
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to Flashlight, Apr 2, 2012
Reply   Reply with quote   Complain
bclaff
Contributing MemberPosts: 672
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to Iliah Borg, Apr 2, 2012

LOL
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Reply   Reply with quote   Complain
bclaff
Contributing MemberPosts: 672
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to Flashlight, Apr 2, 2012

Subtle patterns are unimportant.
Real patterns will really stand out.
The lack of any strong pattern is good news for both cameras.
Some slight pattern noise is not unexpected.
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Reply   Reply with quote   Complain
mlitscher
Regular MemberPosts: 155
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to Flashlight, Apr 2, 2012

Odd, I see "Be sure to drink your Ovaltine."

Reply   Reply with quote   Complain
Lea5
Contributing MemberPosts: 551Gear list
Like?
what does this test mean for my D4?
In reply to bclaff, Apr 2, 2012

is the result good?

bclaff wrote:

Certain types of pattern noise and signal processing can be detected using 2D FFTs (Fast Fourier Transforms).

I performed such an analysis of dark frames (raw) from a D800 and a D4 using ImageJ.
Here's the D800

And the D4

The ISOs from left to right are ISO 50, ISO 1600, ISO 6400, ISO 25600, and ISO 102400 (D4).
The images have been darkened to accentuate the patterns.

Sometimes I find it helpful to view the images from several feet to see subtle features.
The "ideal" FFT would be a uniform gray.
A good guide to interpreting FFTs is here:
http://qsimaging.com/ccd_noise_interpret_ffts.html

Strong vertical or horizontal lines general indicate pattern noise in the horizontal or vertical direction respectively. Both cameras exhibit some pattern noise.

Little "ticks" along an axis is usually a sign of banding, often due to unbalanced channels in a multi-channel readout scheme. This seems very well controlled in both cameras.

The classic "spherical" look of signal processing is absent. There is no evidence of additional signal processing being applied to the raw. In Nikon cameras that general takes place at longer exposures such as 1/4s and longer and will be the subject of a separate study.
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

 Lea5's gear list:Lea5's gear list
Nikon D4 Nikon D800E Nikon AF-S Nikkor 70-200mm f/2.8G ED VR Nikon AF-S Nikkor 14-24mm f/2.8G ED Nikon AF-S Nikkor 24-70mm f/2.8G ED +4 more
Reply   Reply with quote   Complain
voider
Senior MemberPosts: 2,696Gear list
Like?
Re: I see it faintly
In reply to Dave Largent, Apr 2, 2012

It means they are here and they want to give a message

Dave Largent wrote:

bclaff wrote:

In the ISO 102400 I see a sort of square with rounded corners, quite circle-like.

You know, I think you're right. Much more subtle than the usual "sphere".

I think I see it, too. What does it mean?

-- hide signature --
Reply   Reply with quote   Complain
A. Westreich
Senior MemberPosts: 1,195Gear list
Like?
Re: D800 and D4 Pattern Noise Analysis with FFTs
In reply to bclaff, Apr 2, 2012

I think I see dead people.

 A. Westreich's gear list:A. Westreich's gear list
Nikon D7100 Nikon AF Nikkor 35mm f/2D Nikon AF-S DX Micro Nikkor 40mm F2.8 Nikon AF-S Nikkor 80-400mm f/4.5-5.6G ED VR
Reply   Reply with quote   Complain
Marianne Oelund
Veteran MemberPosts: 4,296Gear list
Like?
Clearer presentation
In reply to bclaff, Apr 2, 2012

bclaff wrote:

Certain types of pattern noise and signal processing can be detected using 2D FFTs (Fast Fourier Transforms).

2D FT's suffer the limitation that the output data set is so massive, one can only perceive it readily as an image. Obtaining numerical measurements isn't straightforward.

Row-wise or column-wise 1D FT's can be very revealing of pattern noise, and the data set is small enough to perform further analysis upon, or simple plotting which can show patterns very clearly. One can average a set of rows or columns, then take the DFT (this approach improves SNR for plotting sensor grid-aligned signals), or run DFT on row/columns individually and average the result coefficients (for skewed signals).

D3s average-column DFT, over 240-pixel span for combined Green (G1:G2) channel, ISO 51200 black frame:

Period of each component (in pixels) is 240/frequency. For example, the primary periodicity is 240/40 = 6 pixels. This is the ADC grouping and the DFT is showing mismatch between the 12 ADC channels. Since a specific amplitude is given by the data, it's easy to compare to other cameras.

Now, if Nikon will come through with my D4, I could make this much more topical.

Reply   Reply with quote   Complain
bclaff
Contributing MemberPosts: 672
Like?
Re: Clearer presentation
In reply to Marianne Oelund, Apr 3, 2012

Marianne,

2D FT's suffer the limitation that the output data set is so massive, one can only perceive it readily as an image. Obtaining numerical measurements isn't straightforward.

Row-wise or column-wise 1D FT's can be very revealing of pattern noise,

I agree, but I think there is a place for both since the 2D shows evidence of certain signal processing that 1D doesn't.

In my "spare time" I plan to add 1D to my repertoire

Regards

-- hide signature --
Reply   Reply with quote   Complain
Marianne Oelund
Veteran MemberPosts: 4,296Gear list
Like?
Detecting HPS
In reply to bclaff, Apr 3, 2012

bclaff wrote:

There is no evidence of additional signal processing being applied to the raw. In Nikon cameras that general takes place at longer exposures such as 1/4s and longer and will be the subject of a separate study.

FT's will not show this. Hot-pixel suppression (HPS) produces a constant reduction of the FT across the frequency spectrum.

A histogram of the data, if compared to a non-HPS frame, will clearly show the distortion of the distribution that results.
http://forums.dpreview.com/forums/read.asp?forum=1021&message=40968553

Alternatively, a tabulation of the autocorrelation function (ACF) will immediately show the effect of HPS, which you will see as an elevated coefficient at an image shift of 1 pixel.

For example, this ACF tabulation is for noise that is essentially random. Note that the coefficients for all shifts are small and largely negative, indicating lack of correlation (at n=0 is no shift, which is a reference level for comparison).
n, R, G1, B, G2
0, 98.673, 79.908, 98.839, 117.888
1, -0.685, -0.133, -1.334, 0.136
2, -1.847, -0.57, -2.295, -1.659
3, -0.185, -1.586, -0.824, -0.079
4, -0.94, -0.614, -0.305, -2.314
5, -1.177, 0.034, -2.204, -0.691

The following is from an image that's been processed by HPS. You will see that the coefficients for n=1 are distinctly elevated, indicating correlation between nearest-neighbor pixels, which is what HPS produces:
n, R, G1, B, G2
0, 98.422, 86.965, 98.704, 108.038
1, 8.734, 8.467, 6.106, 5.81
2, 0.504, -0.19, 1.033, 0.789
3, -0.621, -1.049, -2.127, -1.574
4, -2.433, -2.405, -0.914, -3.55
5, -1.006, -1.532, -0.504, -3.137

Reply   Reply with quote   Complain
Iliah Borg
Forum ProPosts: 15,670
Like?
Re: Clearer presentation
In reply to bclaff, Apr 3, 2012

bclaff wrote:
Marianne,

2D FT's suffer the limitation that the output data set is so massive, one can only perceive it readily as an image. Obtaining numerical measurements isn't straightforward.

Row-wise or column-wise 1D FT's can be very revealing of pattern noise,

I agree, but I think there is a place for both since the 2D shows evidence of certain signal processing that 1D doesn't.

Not sure it is just processing. Sensors have interesting structures.

Reply   Reply with quote   Complain
bclaff
Contributing MemberPosts: 672
Like?
Re: Detecting HPS
In reply to Marianne Oelund, Apr 3, 2012

Marianne,

correlation between nearest-neighbor pixels

In my experience this will show in a 2D FFT, but perhaps your test is more sensitive and could be used to form an objective measure.
What are you seeing for the D800 and "HPS"?

Regards
--
Bill (visit me at http://home.comcast.net/~NikonD70/ )

Reply   Reply with quote   Complain
Keyboard shortcuts:
FForum MMy threads