D7000 sensor trivia

Started Jan 29, 2011 | Discussions
Marianne Oelund Veteran Member • Posts: 7,779
D7000 sensor trivia

For basic information about D7000 sensor performance, see my "D7000 Sensible Sensor Data" thread. For potentially interesting, but practically insignificant trivia regarding the sensor and the handling of its data, read on.

After poking around with histograms of D7000 raw data, I've come to a few conclusions:

  1. Sensor maximum sensitivity is ISO 950

  2. Red and blue channel data are being scaled up before recording to the raw file

  3. All channels are scaled up by an additional miniscule factor

  4. Scaling calculations are done carefully, using rounding

A Tale of Missing Codes

All of this comes from analysis, not of the data per se, but of the data that is missing. Given a random data set histogram, we should see a nice normal distribution. Any deviations from that will raise questions about the nature of the data, and I just happen to be one of those persons who can't resist finding answers to such questions.

All of the histograms included here are from bias frames which I took for the D7000 sensor testing. They are created by taking an exposure from an RGB-balanced light source, with the average exposure corresponding to about DN 120. This ensures that the values in the raw file will fall within the 0-255 plot range, giving each DN value its own bin in the plot.
Here is a histogram for ISO 800, which includes all of the color channels:

Although the envelope of the plot shows a good normal distribution, there seems to be a sub-plot intermingled with it. Looking at the diminished values, I realized they represented 25% or 50% missing codes, which suggests that one or two channels are missing data at those values.

The next step was to create separate histograms for each color channel, which then made the situation quite clear.
Red channel:

Green channel:

Blue channel:

Red and Blue Channel Scaling

The green channel histogram has turned out quite solid (except for a single missing code - more about that later). This shows that green data is being transferred to the raw file without any significant manipulation.

The red and blue channels are showing missing codes at very regular intervals (again, with a single additional missing value which I will ignore for the moment). We can use those intervals to infer what scaling factor is being applied. I'll spare you the gory details of calculating that; for the red channel we have 1.126 and for the blue channel, 1.160 (uncertainty is less than 0.2%). In order to find the true data value read from the sensor's converter, divide the red and blue raw-file values by these factors.

The "extra" missing codes

The single extraneous missing codes mentioned above imply that another factor is being applied to the data, which is extremely close to unity. They correspond to a 0.35% increase for green codes, and 0.7% increase for red and blue codes. Only Nikon know why they bothered to apply such fine scalings!

Note that in all cases, the scaling calculations are carried out with extra precision, then rounded; we know this, since the first missing-code interval is only half of the width of the others.

Sensor maximum sensitivity

Up to ISO 800, we are seeing green channel data being used essentially as-is from the sensor. From the results of my sensor testing, it's known that read noise in electron units levels off at ISO 800, remaining constant from there and up. This can infer that higher ISO settings are implemented strictly by digital scaling, so I had a look at histograms from higher-ISO images to see if this could be confirmed.
Here is the green-channel histogram at ISO 1000:

Indeed, we do have digital scaling, and the scale factor is 1.059. This means that when the sensor electronics have been set to their highest gain, the data produced corresponds to ISO 950, and it is then scaled up to produce higher equivalent sensitivities. Examination of histograms from ISO 1100, 1250, 1600, 3200 and 6400, yields the same conclusion.

For the user, the implication of this is that ISO settings above 800 provide no more true image data than ISO 800 itself does; you only lose highlight range in your raw files when you select those settings. Although those high settings have practical value in producing usable JPEG images out of camera, for Raw-only shooting, we are better off staying at ISO 800 and using scaling in post, where we can use tone curve customization to preserve highlights that would have otherwise been clipped.

Other Cameras

Next month, I will be looking at D3 and D3s histograms, which should be interesting. The sensor data suggests that the D3s has an extended analog-gain range relative to the D3, and the histograms should be able to determine this design detail.

Holmes375
Holmes375 Veteran Member • Posts: 3,824
Re: D7000 sensor trivia

Always enjoy your posts even if half your data is over my head
--
Holmes
http://holmes.zenfolio.com/

AndreasE Contributing Member • Posts: 909
Thank you and 2 questions

Thank you Marianne for your effort. Quite interesting.
I do have 2 questions though if you don't mind:

1) Does the spectral distribution of the light source impact the distribution curve and the "behavior" of the sensor electronic to stretch the signal differently in the 3 channels?

2) What level of "rawness" did you analyse? i.e which options in dcraw were turned on in your analysis?

Many thanks,
Andy

OP Marianne Oelund Veteran Member • Posts: 7,779
Re: Thank you and 2 questions

AndreasE wrote:

Thank you Marianne for your effort. Quite interesting.
I do have 2 questions though if you don't mind:

1) Does the spectral distribution of the light source impact the distribution curve and the "behavior" of the sensor electronic to stretch the signal differently in the 3 channels?

Other spectral distributions would only shift the histogram envelopes; the missing code positions would remain fixed where they are. The scale factors are fixed, rather than dependent on white balance settings.

2) What level of "rawness" did you analyse? i.e which options in dcraw were turned on in your analysis?

dcraw -D -W -T -b 256 -g 1 1

Produces an 8-bit TIF containing the lower 8 bits from each 14-bit raw value (clipping any higher values to DN 255).

Many thanks,
Andy

You're welcome, Andy!

leverpastej Forum Member • Posts: 59
Thanks for the analysis

Agrees well with my own findings. I mentioned this in a post a few weeks ago: http://forums.dpreview.com/forums/read.asp?forum=1034&message=37250809

OP Marianne Oelund Veteran Member • Posts: 7,779
Re: Thanks for the analysis

leverpastej wrote:

Agrees well with my own findings. I mentioned this in a post a few weeks ago

Now you know what the exact number is.

Actually, I was a little amused at Nikon's exactness about the digital scaling for higher ISO values. When I went through the histograms at first, I was getting values for the sensor's max ISO that ranged from 924 to 955. Suspecting that the values should be tighter than that, I thought of trying the exact ISO values rather than their nominal ones. That is, ISO 1008 instead of 1000, 1131 instead of 1100 and 1270 instead of 1250 - doing that, the calculated values tightened up to 950 plus/minus 3!
Yet another example of obsession with detail on Nikon's part, I think.

Mannypr Veteran Member • Posts: 3,363
Re: Thanks for the analysis

Sometimes I read a post that makes me feel an ignorant clown :(. I am humbled by your knowledge on the subject. Apart from this just one question which I had to rethink a couple of times before posting it because it might make me look like a fool but here goes...Does all this mean that from Iso 800 downwards all amplification of the electrical signal is done in the analog domain and from 800 upwards it's done digitally ?

PS...Thank you for going into all the work iinvolved.

 Mannypr's gear list:Mannypr's gear list
Canon PowerShot A640 Nikon D90 Nikon D7000 Nikon AF-S DX Nikkor 18-55mm f/3.5-5.6G II Nikon AF-S DX Nikkor 18-105mm F3.5-5.6G ED VR +8 more
stuntmonkey
stuntmonkey Senior Member • Posts: 2,735
Re: D7000 sensor trivia

Marianne Oelund wrote:

  1. Sensor maximum sensitivity is ISO 950

This begs the question... what was the cutoff point for the D90/D300?

lovEU Veteran Member • Posts: 3,135
Thanks ...

… for having done all the work and sharing your results here.

Marianne Oelund wrote:

  1. All channels are scaled up by an additional miniscule factor

Has this to do with that Nikon “ISO fine tuning”, in the sense you mentioned their obsession with detail?

For the user, the implication of this is that ISO settings above 800 provide no more true image data than ISO 800 itself does; you only lose highlight range in your raw files when you select those settings. Although those high settings have practical value in producing usable JPEG images out of camera, for Raw-only shooting, we are better off staying at ISO 800 and using scaling in post, where we can use tone curve customization to preserve highlights that would have otherwise been clipped.

I think that’s the reason Iliah suggested on several occasions shooting with a camera specific “sweet spot” ISO and “underexpose” available light scenes instead of using high ISO values. In many cases one will need a raw converter being capable of using such data then. RPP is great in this regard (as well as in other respects).
--
regards, eric

Bob in Baltimore
Bob in Baltimore Senior Member • Posts: 1,000
Re: D7000 sensor trivia

Thanx for the great analysis Marianne! As a raw shooter. my take away message is don't shoot over ISO 800. Handle it in PP. Thanx!

It may interest you and others that we have observed a related issue in MODIS, a space-borne imaging radiometer for studying the Earth's environment. When one of our university research teams made histograms from a large number of images covering the entire Earth, they found jagged histograms. In that case, the 14-bit analog-to-digital converter (ADC) was not filling the bins equally as the result of non-linear response. It's very difficult to manufacture them to the tightest tolerances. In this case, the alternate set of electronics was much better, and we switched to that side for the best radiometric precision.

Our parts were screened, and the two best units were selected. Imagine selecting ADC's for a production run. Some will be much better than others. It will be the luck of the draw. But, like your example, it will likely have no practical impact on our images.
--
Bob in Baltimore

 Bob in Baltimore's gear list:Bob in Baltimore's gear list
Nikon AF-S Nikkor 70-300mm f/4.5-5.6G VR Nikon AF-S Micro-Nikkor 105mm F2.8G IF-ED VR Nikon AF-S Nikkor 85mm f/1.4G Nikon AF-S Nikkor 35mm f/1.4G Nikon AF-S Nikkor 24-85mm F3.5-4.5G ED VR
leverpastej Forum Member • Posts: 59
Re: Thanks for the analysis

Marianne Oelund wrote:

leverpastej wrote:

Agrees well with my own findings. I mentioned this in a post a few weeks ago

Now you know what the exact number is.

Yes, nice to know.

Actually, I was a little amused at Nikon's exactness about the digital scaling for higher ISO values. When I went through the histograms at first, I was getting values for the sensor's max ISO that ranged from 924 to 955. Suspecting that the values should be tighter than that, I thought of trying the exact ISO values rather than their nominal ones. That is, ISO 1008 instead of 1000, 1131 instead of 1100 and 1270 instead of 1250 - doing that, the calculated values tightened up to 950 plus/minus 3!
Yet another example of obsession with detail on Nikon's part, I think.

Personally I can only feel sorry that Nikon fiddles with the raw data this way. Seems utterly pointless to destroy highlights only to insert gaps in the histogram. Sure, we can trick the camera to do the right thing by never going above ISO 800 (or 1000), instead underexposing + push processing to wanted ISO, but we shouldn't have to. Perhaps Nikons obsession with detail, as you put it, has made them completely miss the big picture here. We don't need digital scaling of raw data. Just get rid of it. I also don't like how Nikon subtracts the black level offset in the raw data, clipping negative values to zero.

Ray Soares Veteran Member • Posts: 3,121
Re: D7000 sensor trivia

"you only lose highlight range in your raw files when you select those settings."
Great info!
Thanks!
--
Ray Soares

See my pictures at http://www.pbase.com/raysoares

 Ray Soares's gear list:Ray Soares's gear list
Nikon D7000 Nikon D4 Nikon AF-S Nikkor 70-200mm f/2.8G ED VR II Nikon AF-S Nikkor 24mm f/1.4G ED Nikon AF-S Nikkor 85mm f/1.4G +1 more
Swedish Hambern Contributing Member • Posts: 751
Auto ISO <800 then?

Seems like setting Auto ISO maximum to 800 should be quite wise then.

Mannypr Veteran Member • Posts: 3,363
Re: Thanks for the analysis

Mannypr wrote:

Sometimes I read a post that makes me feel an ignorant clown :(. I am humbled by your knowledge on the subject. Apart from this just one question which I had to rethink a couple of times before posting it because it might make me look like a fool but here goes...Does all this mean that from Iso 800 downwards all amplification of the electrical signal is done in the analog domain and from 800 upwards it's done digitally ?

PS...Thank you for going into all the work iinvolved.

My question still stands...can anybody answer it for me ?

 Mannypr's gear list:Mannypr's gear list
Canon PowerShot A640 Nikon D90 Nikon D7000 Nikon AF-S DX Nikkor 18-55mm f/3.5-5.6G II Nikon AF-S DX Nikkor 18-105mm F3.5-5.6G ED VR +8 more
Ray Soares Veteran Member • Posts: 3,121
Re: Auto ISO <800 then?

Only if you shoot raw.

The problem with auto-iso is that the camera, after reached maximum ISO, will go to a slower speed than the minimum you've stablished!
Lets test it...
Best
--
Ray Soares

See my pictures at http://www.pbase.com/raysoares

 Ray Soares's gear list:Ray Soares's gear list
Nikon D7000 Nikon D4 Nikon AF-S Nikkor 70-200mm f/2.8G ED VR II Nikon AF-S Nikkor 24mm f/1.4G ED Nikon AF-S Nikkor 85mm f/1.4G +1 more
digitalnoise Contributing Member • Posts: 868
ADL bracketing?

If the sensor maxes out just above 800iso, and given the oft mentioned propensity to overexpose at times, does the wisdom of nikon's adl (which I had personally dismissed) finally come to light?

geo444
geo444 Contributing Member • Posts: 502
Re: Thanks for the analysis

Mannypr wrote:

...Does all this mean that from Iso 800 downwards all amplification of the electrical signal is done in the analog domain and from 800 upwards it's done digitally ?

this was the case for Eos 20d, 5d... except the threshold was 1600 rather than 800 !

seems there is No Analog Amp. in the New Sony Sensors (Nikon D7k, Pentax K5)

... what's new is the Digital CDS, canceling Conversion Noise just after the ADC !
http://www.chassimages.com/forum/index.php/topic,112598.0.html
... just more Digital Precessing from 800 Iso upwards ?
CDS = Correlated Double Sampling

geo444
geo444 Contributing Member • Posts: 502
Re: Thanks for the analysis

Mannypr wrote:

...Does all this mean that from Iso 800 downwards all amplification of the electrical signal is done in the analog domain and from 800 upwards it's done digitally ?

this was the case for Eos 20d, 5d... except the threshold was 1600 rather than 800 !

seems there is No Analog Amp. in the New Sony Sensors (Nikon D7000, Pentax K-5)
what's new is the Digital CDS, canceling Conversion Noise just after the ADC !
http://www.sony.net/SonyInfo/technology/technology/theme/cmos_02.html
... just more Digital Precessing from Iso 800 upwards ?

CDS = Correlated Double Sampling

OP Marianne Oelund Veteran Member • Posts: 7,779
Gain split

Mannypr wrote:

Does all this mean that from Iso 800 downwards all amplification of the electrical signal is done in the analog domain and from 800 upwards it's done digitally ?

You understand the basic concept. From ISO 1000 and up, the sensor stays at its maximum gain (equivalent to ISO 950) and the sensor's digital output is numerically scaled to the ISO setting. That is, for all ISO settings in the 1000 to 25600 range, the sensor is doing exactly the same thing.

OP Marianne Oelund Veteran Member • Posts: 7,779
You have to love marketing spins

geo444 wrote:

... what's new is the Digital CDS, canceling Conversion Noise just after the ADC !
http://www.chassimages.com/forum/index.php/topic,112598.0.html
... just more Digital Precessing from 800 Iso upwards ?

They'd have us believe all noise magically disappears - see the note on the output in figure 2 which says, "It is Digital. Therefore it is noise free."

In reality, they are only canceling reset noise, which is a fixed pattern noise that sensor testers ignore anyhow.

The paper discloses very little detail that's useful for engineering analysis. Figure 2 doesn't even have enough detail to show how the control signals at the top of figure 3 are being used in the circuitry.

Keyboard shortcuts:
FForum MMy threads