Canon 1D4 sensor data

[Roger Clark]

Hi,

I'm wondering if anyone has read anything from Canon regarding the 1D4 sensitivity of the sensor. Canon indicated the sensor in the 7D was more sensitive and the sensor data shows that to be the case; results here if you haven't seen them: http://www.clarkvision.com/articles/...mance.summary/

I'm analyzing data for the 1D4 now and the numbers are coming out impressively high, indicating Canon has done another fundamental improvement. These are preliminary numbers, but the full well = 71,000 and the high ISO read noise hits a low of 2 electrons at ISO 12,800 (this is a new low; the 5D2 is 2.5 electrons). The 71,000 value is at a level similar to the 1DIII with is 7.2 micron pixels but with the smaller 1D4's 5.7 micron pixels, for a 60% improvement, and with lower noise. This would put the AIQ on figure 9 on the above web page well above the predicted point and on par with the 5D2. Very impressive if correct!

I should have the full analysis up in a few days. So again, I'm looking for confirmation from Canon that they have improved the sensitivity, for example, is there a 1D4 white paper yet?

Roger

[Steve Canuel]

Roger,
Is this what you're looking for?
http://www.usa.canon.com/dlc/control...rticleTypeID=2

Appreciate you taking the time to do this kind of stuff.

[Axel Hildebrandt]

Thanks, Roger. Sounds very good, here is the direct link: http://www.usa.canon.com/uploadedima...20IV%20WP1.pdf

On page 25 it says in regard to high ISO performance, the Mark IV is almost 2 stops better than the Mark III due to the processor and sensor, on page 40 it says that the sensitivity of the sensor is 1 stop better over the Mark III.

[Emil Martinec]

How's the pattern noise? Also, are you testing intermediate ISO's? I'm curious if they're still using the two-stage hardware gain they used in previous 1 series and 5D, as opposed to software gain used in xxD, 7D and 5D2.

[Charles Glatzer]

Non-scientific of course, but In two weeks of using the camera a field I have noticed a gain of about 1 EV plus in ISO noise reduction and .5 EV in highlight headroom over the Mark III when using LR2 to convert both files. BTW- the MIV AWB appears more accurate than with any other digital camera I have used.

At this point I still find the overall IQ better with my IDs M III bodies with lower ISO use, but the quick and accurate AF, high burst and buffer rate of the MIV are a welcomed addition when necessary.

Chas

[Roger Clark]

Steve, Axel,
Thanks. That is exactly what I needed. Canon describes doing several things resulting in a stop of improvement in system sensitivity, and it looks like they have achieved that. That. combined with better fill factors and gapless micro lenses, means if they hadn't improved system sensitivity, the 1D4 would still be on par with sensors with larger pixels. So all this combined is an impressive improvement. This will be great for very low light and high ISO work and will set a new performance standard.

Emil Peter H supplied the data and included some intermediate ISOs but I have not analyzed them yet. Would you like the data? I'm not sure I'll have time short term to analyze the intermediate ISOs. I've got a major update on my sensor performance page in the works and now I'll have to add a new model for the 1D4. This is the first camera that really breaks ahead of my model predictions, indicating a significant improvement in system sensitivity. The 7D is also better than my model predictions, but only by about 10%.

I'll evaluate the pattern noise too but it looks really low. I did a better fit to the data and derived a gain at ISO 100 of 5.95 electrons/DN and read noise is 2.5 electrons. not 2, at iso 12,800. I may refine these results a little more before the web pages go up, but Peter's data are very tight (great job Peter).

Roger

[Roger Clark]

OK, an update. Some of the numbers I was getting just seemed off, but I just couldn't find an error in my analysis (I've mostly automated the procedure with custom software to minimize the errors). I found a problem in that my less than 2-month-old version of DCRAW was not quite reading the 1D IV raw files correctly. The generated tiffs were readable by photoshop but seemed strangely smooth, and ImagesPlus could not read the tiff files correctly, but my analysis software did read the files OK. I updated DCRAW and things are much more consistent. The full well and gains came down a little, but still very impressive.

The web page where results will go is at:
http://www.clarkvision.com/articles/...on-canon-1div/

I put up an ISO 12800 image, heavily stretched for Emil to show the amazing uniformity in the read noise. The ISO 100 gain is now 4.2, still preliminary, but I am more confident in the analysis now.

I'll go away now and complete it....
Roger

[Emil Martinec]

I put up an ISO 12800 image, heavily stretched for Emil to show the amazing uniformity in the read noise. The ISO 100 gain is now 4.2, still preliminary, but I am more confident in the analysis now.

Roger

4.2 for the gain sounds much more reasonable. Perhaps you could email me a link to a zip archive of the raw files, and I might do some checks if I can find the time.

Thanks for the crop...

[Charles Glatzer]

Roger,

You mention the sensor being capable of 15 stops, but limited to just over 11 stops in your data, but could you tell me based on your analysis the dynamic range of the Mark IV in f/stops in actual field use from black to white clipping. I am finding approx 3.5 f/stops above the midpoint before highlight clipping occurs. I have not had a chance to determine black clipping values. Non-linear conversion. Much appreciated.

Best,

Chas

[Axel Hildebrandt]

Roger, I just read the updated website about the 1D4. Do you have the values of the Mark III somewhere for comparison?

Thanks!

[Roger Clark]

The dynamic range in actual use is given in Table 1 as a fuction of ISO. At least as defined as max signal at a given ISO divided by the noise floor at that ISO. The dynamic range is limited by electronics at low ISOs and at high ISOs is slightly better than the 5D Mark II.

I have the full sensor analysis on the web page now, just refining a few things. I might make a few new plots. It is really quite impressive how the model tracks the data to within a couple of percent over 11 stops. And that agreement proves how the sensor noise is dominated photon noise over most of its range (that is a good thing).

Roger

[John Chardine]

FWIW, here's one from today with my mk IV at ISO 3200. I had to use this high ISO because of low light as the sun was setting.

As I am of the opinion that ISO and images should be judged after processing, this one has been, processed that is- high ISO NR was set to Standard in camera and the image was processed with DPP using these settings. Normal NR performed in PP on BG only; simple stuff, not rocket science. The take home message for me is that ISO 3200 is usable. I am in shock.

Canon EOS-1D Mark IV, 500/f x 1.4 tc
capture date: 13 February 2010, 4:38 PM
exposure program: Aperture Priority
ISO speed: 3200
shutter speed: 1/1600
aperture: f5.6
exposure bias: +0.0
metering: Pattern
flash: OFF

[Charles Glatzer]

The dynamic range in actual use is given in Table 1 as a fuction of ISO. At least as defined as max signal at a given ISO divided by the noise floor at that ISO. The dynamic range is limited by electronics at low ISOs and at high ISOs is slightly better than the 5D Mark II.

I have the full sensor analysis on the web page now, just refining a few things. I might make a few new plots. It is really quite impressive how the model tracks the data to within a couple of percent over 11 stops. And that agreement proves how the sensor noise is dominated photon noise over most of its range (that is a good thing).

Roger

Roger,

I know the science says 11 stops. But, assuming you go from a mid-tone value I do not see this in real world practice via histogram in-camera or when RAW processing. What I do see is approx a 7 stop range with the Mark IV clipping highlights approx 3.5 EV above midtone. I am more concerned with practical application in the field. If you could help me better understand this would it be appreciated.

Best and Thanks,

Chas

[Roger Clark]

Roger,

I know the science says 11 stops. But, assuming you go from a mid-tone value I do not see this in real world practice via histogram in-camera or when RAW processing. What I do see is approx a 7 stop range with the Mark IV clipping highlights approx 3.5 EV above midtone. I am more concerned with practical application in the field. If you could help me better understand this would it be appreciated.

Best and Thanks,

Chas

Chas,
Your experience looks quite consistent with the sensor data. The 11 stops is from clipping to noise floor. Depending on the metering, 18% gray is 2.5 stops down from clipping, so you would then have 11-2.5 = 8.5 stops below 18% gray. Most metering in my experience is a little conservative, so move down another stop and then you have 7.5 stops below 18% gray to 3.5 stops above, close to your experience. It seems that the clipping point on digital cameras (the point where the "blinkies" turn on, is pretty conservative and one really has more headroom in the raw data. Further, often one can recover highlights (as you probably know) somewhat above that too. That is because the signals in the red and blue channels are lower than in the green channel so they don't saturate as soon.
Does this answer your question?

Roger

[Emil Martinec]

Chas,

As Roger mentioned, the DR he reports is the engineering definition of this quantity -- the max recordable signal divided by the noise with no signal. This very much an upper bound; there isn't much utility to the deepest shadow part of the range, where signal/noise is very low (where signal=average tonal level). I would expect the S/N of the 1D4 to be rather similar to the 1D3 qualitatively. Here is the 1D3 S/N as a function of signal at various ISO, from DxO:



For S/N in stops, divide the vertical axis by six; the bottom end of 7 stops of DR on the 1D3 at ISO 100 occurs a little below 1% (2^7=128, and 1/128~.8%) where the S/N is 24dB or 4 stops, or 16:1. So this is your minimum acceptable S/N, and leaves another 4 stops of range below that where the camera can record data but it is below your standards for quality.

[Charles Glatzer]

Roger and Emil,

Thanks for the usable exposure conformation.

I greatly appreciate the technical insight, further explanation, and clarification.

Warmest Regards to you both,

Chas

[Cecil Kirksey]

Chas,

As Roger mentioned, the DR he reports is the engineering definition of this quantity -- the max recordable signal divided by the noise with no signal. This very much an upper bound; there isn't much utility to the deepest shadow part of the range, where signal/noise is very low (where signal=average tonal level). I would expect the S/N of the 1D4 to be rather similar to the 1D3 qualitatively. Here is the 1D3 S/N as a function of signal at various ISO, from DxO:



For S/N in stops, divide the vertical axis by six; the bottom end of 7 stops of DR on the 1D3 at ISO 100 occurs a little below 1% (2^7=128, and 1/128~.8%) where the S/N is 24dB or 4 stops, or 16:1. So this is your minimum acceptable S/N, and leaves another 4 stops of range below that where the camera can record data but it is below your standards for quality.

So a SNR of 16 (linear) is required for acceptable IQ? The normal definition of dynamic range as used in camera analysis does not make much sense since as you suggest an acceptable minimum SNR is required to have acceptable IQ. But acceptable IQ seems to be photographer dependent. Basically the maximum dynamic range given a meaningful definition of acceptable IQ will require a certain SNR which will be photon shot noise limited and therefore depend on maximum well capacity and not so much on read noise.

[Roger Clark]

So a SNR of 16 (linear) is required for acceptable IQ? The normal definition of dynamic range as used in camera analysis does not make much sense since as you suggest an acceptable minimum SNR is required to have acceptable IQ. But acceptable IQ seems to be photographer dependent. Basically the maximum dynamic range given a meaningful definition of acceptable IQ will require a certain SNR which will be photon shot noise limited and therefore depend on maximum well capacity and not so much on read noise.

Not only photographer dependent, but time dependent. Film, for example. does not have a very high S/N, yet people seemed happy with it. Then when digital came along, some people didn't like digital because they couldn't see grain (noise)! Film never gets above a S/N of about 60 to 80, and that is in the highlights. In middle gray, it is typically below 20, and just gets worse from there. So with film, many protions of images had S/N < about 1, but suddenly that is no longer acceptable. For example, this page shows that a 1DII has higher S/N in middle gray and lower at ISO 1600 than Fuji Velvia ISO 50:
http://www.clarkvision.com/articles/...gnal.to.noise/
(the portion less than 10,000 on the horizontal axis on Figure 1).

Also, when you have lots of pixels (which we do now with 12+ megapixel cameras), your eye integrates signal and you can see image detail well below S/N = 1. In fact, you can see image detail below S/N <0.1, just not fine detail. There are many such examples on this page:
http://www.clarkvision.com/articles/...t.photography/
The S/N per pixel is listed for many examples. Clearly S/n < few is not acceptable for beautiful pleasing images, but does have a place in things like science and surveillance photography, or to show an interesting effect, like your hand making luminescence in one of the bio-bays in Puerto Rico (I did that with a 5DII and 20 mm f/2.8 lens).

Roger

[Emil Martinec]

So a SNR of 16 (linear) is required for acceptable IQ? The normal definition of dynamic range as used in camera analysis does not make much sense since as you suggest an acceptable minimum SNR is required to have acceptable IQ. But acceptable IQ seems to be photographer dependent. Basically the maximum dynamic range given a meaningful definition of acceptable IQ will require a certain SNR which will be photon shot noise limited and therefore depend on maximum well capacity and not so much on read noise.

You will note that I was careful to state that the SNR 16 was Chas' criterion based on his statement that he thought the (usable) range of the 1D4 was about 7 stops. Different people have different standards.

The engineering definition used in camera analysis is fine, so long as you know what it is telling you; and it has the advantage of being an objective, quantitative standard. One simply has to realize that the useful range will be less, and how much less depends on how picky you are.

As to whether shot noise is the determining factor, one can answer that based on the curves. When the SNR is shot noise limited, the SNR rises 3dB for every stop increase in signal (doubling of linear tonal level of the raw data). When the read noise comes into play, the SNR curve bends down and starts to drop 6dB per stop when totally read noise dominated; this is why the curve is steeper at low signal. So one can see whether read noise is an issue at some particular ISO if for your particular standard for minimum acceptable SNR, the curve has steepened into the read noise portion by that point.

For comparison purposes, here is the Canon 1D3 again:


and here is the Nikon D3x:


Notice how much flatter the D3x curve is at the lower reaches; the Sony Exmor sensor it uses has a massively parallel readout (one ADC per sensor column), with much better low ISO DR. It doesn't matter so much if your SNR criterion is 16:1 or more, but if it's 8:1 or better the D3x has as much as a stop more of low ISO DR per pixel (not that I think that the per pixel criterion is necessarily the right one; it depends on the application).

Finally, one thing that these sorts of curves don't show is the strength of pattern noise. Pattern noise is only a small contributor to SNR, but since it is coherent over many pixels the eye picks it up much more easily than spatially random noise. Nikon has largely solved this issue, while Canon seems content to have deep shadows looking like burlap, limiting the useful DR.

[Roger Clark]

Finally, one thing that these sorts of curves don't show is the strength of pattern noise. Pattern noise is only a small contributor to SNR, but since it is coherent over many pixels the eye picks it up much more easily than spatially random noise. Nikon has largely solved this issue, while Canon seems content to have deep shadows looking like burlap, limiting the useful DR.

Emil,

The DXO curves you posted look fine for the Nikon, but are definitely not representative of the Canon. The Canon low iso (top curve) shows curvature over the entire range, while the Nikon is straight at the top 2/3rds or so. The only conclusion I can draw from that is that DXO does not appear to use actual raw data. (Yes, they may use a data from a raw file, but they must put it through conversion software which changes the raw data. I've seen numbers out of their tests that are quite different from true raw data results. So their curves and results are a combination of sensor performance and raw conversion software. That makes it very hard to understand true trends.

Here is log S/N versus exposure in stops for a 1D2 and Canon S70 P&S, Figure 3b at:
http://www.clarkvision.com/articles/...l.size.matter/

Note all the curves are straight at the top end. I've never seen a camera that wasn't.

Regarding Nikon versus Canon fixed pattern noise, Nikon cooks their raw data so it isn't true raw, even (in some/all? cases) applying a median filter and, as you know, cutting off the bottom end. On astro images, Nikon's median filter deletes faint stars like they are noise, at least on some cameras. So I don't think Nikon solved the issue, except to hide it at the expense of some actual image data.

The fixed pattern noise is/has been a bane of the technology for many years but is getting better each year (you think it's bad in DSLRs, it's much worse in infrared sensors). Canon has had leading lows (meaning better) in read noise technology, but that unfortunately makes the fixed pattern noise all the more apparent, even after they reduce it with each generation. Now that we have read noise below 2 electrons in the 1D4, with fixed pattern noise still apparent at more than 10x smaller, that means fixed pattern noise must be kept within 0.2 electrons on average! If for example, one row latches higher or lower by even 1 electron for a few pixels. it would stick out like a sore thumb. We are not even seeing that in the 1D4 at high ISO. Impressive! How do they do that at room temperature?

Roger

[Roger Clark]

I just computed the range in the noise image on my 1D4 web page: it is only 1.7 electrons from black to white! Perhaps this is why the histogram in Figure 1 is lumpy: we are seeing the quantization of electrons.

http://www.clarkvision.com/articles/...on-canon-1div/

Roger

Edit: I put a scale in electrons on the histogram and we are not seeing electron quantization. That histogram is logarithmic from ImagesPlus (I forgot about that) so the side lobes are some low level noise that does not appear in a linear photoshop histogram (which I also added to the figure). The side lobes may represent some low level fixed pattern noise. And the range one the image is 7 electrons, not 1.7.

Roger

[Emil Martinec]

Emil,

The DXO curves you posted look fine for the Nikon, but are definitely not representative of the Canon. The Canon low iso (top curve) shows curvature over the entire range, while the Nikon is straight at the top 2/3rds or so. The only conclusion I can draw from that is that DXO does not appear to use actual raw data. (Yes, they may use a data from a raw file, but they must put it through conversion software which changes the raw data. I've seen numbers out of their tests that are quite different from true raw data results. So their curves and results are a combination of sensor performance and raw conversion software. That makes it very hard to understand true trends.

Here is log S/N versus exposure in stops for a 1D2 and Canon S70 P&S, Figure 3b at:
http://www.clarkvision.com/articles/...l.size.matter/
Note all the curves are straight at the top end. I've never seen a camera that wasn't.

The DxO data is not measured on demosaiced output, if that is what you mean by converted. The SNR curves are as I understand fits to a model of noise that includes read noise, shot noise, and pixel response non-uniformity (PRNU); you can find details at their website, which has some documentation of what their methodology is. So the curves they plot are not the direct measurements of the raw data but the best fits to the model using measurements directly on raw data. There are some glaring examples that are way off (the LX3 comes to mind) but I believe their results for the 1D3 and most cameras of interest to people here are accurate; they compare well with my own analysis.

I think a big difference from the way you analyze data is that they are measuring directly pixel standard deviations rather than using difference images. The latter for instance removes the effect of PRNU which the former incorporates. At the lowest ISO, the effect of PRNU can be quite substantial and I believe it accounts for the substantial curvature of the 1D3 plots at the upper end at low ISO; see the mouseover in figure 7 at
http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/
for an analysis I did on my 20D of the strength of PRNU at base ISO (I used difference images to fit the read/shot noise, then subtracted that from single frames to estimate the PRNU).

Much of what you are seeing in the discrepancy between the two graphs (1D3 and D3x) at the upper end is what I mentioned above, that Nikon has achieved much better control over pattern noises; that includes not just fixed amplitude effects that are buried below the read noise, it also includes variations in per column gain (something that is atrocious in the 7D, for instance).

Regarding Nikon versus Canon fixed pattern noise, Nikon cooks their raw data so it isn't true raw, even (in some/all? cases) applying a median filter and, as you know, cutting off the bottom end. On astro images, Nikon's median filter deletes faint stars like they are noise, at least on some cameras. So I don't think Nikon solved the issue, except to hide it at the expense of some actual image data.

As far as I am aware, the only situation where Nikon filters their raw data is on long exposures (the threshold where that begins is camera dependent, for the D300 and D3 it's 1/4 sec). And yes that is bad news for astrophotography, but not for the faster shutter speeds relevant to daytime nature photography. Yes they clip blacks, again that is only relevant when there is nearly no ambient light, and you want to dig down into the raw data near zero signal. Black clipping has no effect on well exposed daylight images.

The fixed pattern noise is/has been a bane of the technology for many years but is getting better each year (you think it's bad in DSLRs, it's much worse in infrared sensors). Canon has had leading lows (meaning better) in read noise technology, but that unfortunately makes the fixed pattern noise all the more apparent, even after they reduce it with each generation. Now that we have read noise below 2 electrons in the 1D4, with fixed pattern noise still apparent at more than 10x smaller, that means fixed pattern noise must be kept within 0.2 electrons on average! If for example, one row latches higher or lower by even 1 electron for a few pixels. it would stick out like a sore thumb. We are not even seeing that in the 1D4 at high ISO. Impressive! How do they do that at room temperature?
Roger

I've heard that Nikon has dropped the high ISO read noise of the D3s to 2.5 electrons (from around 5 on the D3), but seems to have controlled the pattern noise even better as one can see from the insanely high ISO images coming from it. I have no idea how they are achieving that, it's just astonishing. I hope Canon is able to catch up; my sense is that they've gotten better at pattern noise in the readout, but have allowed the PRNU to get worse (at least in the 7D, I haven't looked at the 1D4).

[Roger Clark]

Emil,
Quick note, you are comparing Canon 20D noise to Nikon D3 noise. Those are different by several generations of technology.

Roger

[Emil Martinec]

Emil,
Quick note, you are comparing Canon 20D noise to Nikon D3 noise. Those are different by several generations of technology.

Roger

No, I am saying that PRNU can be a substantial contributor to noise at low ISO, albeit at high signal levels where SNR is quite high. I pointed to graphs from a detailed analysis of a 20D, because that is what I have detailed data for. I did a quick check of my 1D3 some time ago, and saw that single frame noise at high signal is higher than the std dev results from difference frames at base ISO, at high signal, consistent with the 1D3 results from DxO. I did a lot of analysis of the 7D when it first came out; it has a huge amount of gain variance from column to column, to the extent that it screwed up the demosaic results from ACR 5.5 (that was beta support, but the point is that the raw data was sufficiently damaged that using the same algorithm as for other Canon DSLRs didn't work well).

PRNU affects the SNR graphs in precisely the way that is observed in the DxO plot for the 1D3; since it is gain variance issue, it provides a strict cap on the maximum SNR that can be achieved.

Since you have the 1D4 data, you can check for yourself if you're interested as to whether it's a significant contributor to bending down the low ISO pixel std dev when you don't eliminate it by taking difference frames.

[Roger Clark]

Emil,

I remain unconvinced. The reasons are as follows. I test sensors a lot (professionally) and deriving the top end S/N is very difficult. With DSLRs, we are talking S/N at low ISOs above 200. That means that the uniformity of the light on the sensor must be uniform to much better than that level, and that is extremely difficult to do (e.g. at least to 1 part in 1000). If one uses a single source, you have 1/r squared effects, then if using optics, light falloff effects, then there are reflections and flare. If one is using extended sources and no optics you still have reflections which are very difficult to control.

I looked at some of my data and when stretched, yes I can see fixed pattern noise in the 1D4 data, but it is extremely small. I did some standard deviations on single 200x200 pixel images and they were ~2x higher than
the subtraction method. I then investigated and found what I expected: ramps due to uneven lighting. And this is only 200 pixel square areas. I did partial corrections for the ramp and reduced the difference to 20%, and it looks like with a good model for the light distribution, the single image standard deviation would be about 10% higher than two subtracted images/root 2. So fixed pattern noise would be very small and would not be visible on a log plot.

As the light levels drop, the S/N drops and the ramps in light level become less significant, so the difference in single versus two subtracted images becomes less, unless there is low level fixed pattern noise.

Regarding the 7D, what you observed are side effects due to the two green filters being used in the Bayer filter array, and software that didn't deal with it properly.

Another factor in fixed pattern noise and single image analysis, is one must separate the Bayer filters or the different transmission of the red, green, and blue filters, or those transmission differences will contribute to the derived standard deviation. I did separate the filters for my analysis.

Roger

[Emil Martinec]

Emil,

I remain unconvinced. The reasons are as follows. I test sensors a lot (professionally) and deriving the top end S/N is very difficult. With DSLRs, we are talking S/N at low ISOs above 200. That means that the uniformity of the light on the sensor must be uniform to much better than that level, and that is extremely difficult to do (e.g. at least to 1 part in 1000). If one uses a single source, you have 1/r squared effects, then if using optics, light falloff effects, then there are reflections and flare. If one is using extended sources and no optics you still have reflections which are very difficult to control.

I looked at some of my data and when stretched, yes I can see fixed pattern noise in the 1D4 data, but it is extremely small. I did some standard deviations on single 200x200 pixel images and they were ~2x higher than
the subtraction method. I then investigated and found what I expected: ramps due to uneven lighting. And this is only 200 pixel square areas. I did partial corrections for the ramp and reduced the difference to 20%, and it looks like with a good model for the light distribution, the single image standard deviation would be about 10% higher than two subtracted images/root 2. So fixed pattern noise would be very small and would not be visible on a log plot.

As the light levels drop, the S/N drops and the ramps in light level become less significant, so the difference in single versus two subtracted images becomes less, unless there is low level fixed pattern noise.

Regarding the 7D, what you observed are side effects due to the two green filters being used in the Bayer filter array, and software that didn't deal with it properly.

Another factor in fixed pattern noise and single image analysis, is one must separate the Bayer filters or the different transmission of the red, green, and blue filters, or those transmission differences will contribute to the derived standard deviation. I did separate the filters for my analysis.

Roger

I used to examine std devs over quite a few patches much smaller than 200x200, precisely to avoid gradients. I can't remember if I did so to test my 20D, but my current method for checking PRNU involves subtracting the two green subarrays and taking the std dev of that. DxO tests the bodies without a lens, IIRC. I don't think gradients would explain the discrepancy between their 1D3 and D3x results.

As for the 7D, here is the difference of two successive frames at the same exposure, G1 channel (I too separate the Bayer array)


and here is what one gets when one subtracts the G1 subarray from the G2 subarray of one of these images


Subtracting the two green subarrays is a good way of removing the effects of gradients without removing the effects of PRNU. For the above example, PRNU (mostly column gain variation) is at least as large as shot noise. This is a midtone at base ISO, BTW.

[Roger Clark]

Subtracting the two green subarrays is a good way of removing the effects of gradients without removing the effects of PRNU.

I disagree. If each line and filter has a different gain, then subtracting two different filters could actually increase the fixed pattern noise.

Roger

[Cecil Kirksey]

So what is a good number for PRNU? 1%? .3%? Hypo: 90,000 electron well capacity (one-to-one electron-to-photon) yields shot noise of 300 electrons. A 1% PRNU yields a noise of 900 electrons assuming that the PRNU is Gaussian with sigma of 1%. A PRNU of .3% still yields a sigma of 270 elctrons. So it looks like the maximum SNR could easily be PRNU limited. So how would a manufacturer control this? Measure each photosite with a colored laser and calculate an offset gain? HMMM!!!

[Emil Martinec]

I disagree. If each line and filter has a different gain, then subtracting two different filters could actually increase the fixed pattern noise.

Roger

Um, that's the point. Nonuniform illumination of the sensor is eliminated (or rather, rendered an irrelevant effect several orders of magnitude smaller than other effects); what remains are effects coming from variations in the transmissivity of color filters, conversion efficiency of the photosites, pixel/column gain, etc. In other words, non-uniformity of pixel response to incoming light. One also has of course shot noise and read noise as well (times sqrt[2]), but those can be separately measured from difference frames of the same channel and subtracted out to isolate the PRNU.

[Roger Clark]

My point was you don't know the magnitude of the fixed pattern noise. For example, by subtracting two different filters, you don't know if you have increase it by 2x or completely suppressed it. It seems to me that a better way is to model the low frequency background in single images and ratio it out, leaving the shot noise and fixed pattern noise.

Roger

[Roger Clark]

So what is a good number for PRNU? 1%? .3%? Hypo: 90,000 electron well capacity (one-to-one electron-to-photon) yields shot noise of 300 electrons. A 1% PRNU yields a noise of 900 electrons assuming that the PRNU is Gaussian with sigma of 1%. A PRNU of .3% still yields a sigma of 270 elctrons. So it looks like the maximum SNR could easily be PRNU limited. So how would a manufacturer control this? Measure each photosite with a colored laser and calculate an offset gain? HMMM!!!

Well, first I would ask how many images have you seen posted or processed yourself where pixel response non-uniformity (PRNU) is an issue? (Ignore 7D raw converters that didn't properly account for the new two different green filters.) I've not seen it.

But if you do have that problem, and your camera is stable, there isa simple solution. It is called a flat field and has been used for decades to calibrate sensors. You measure a uniform scene, like a blank wall and use it to ratio out non-uniformities, including light fall-off.

Personally, I don't see PRNU as an issue and have never had to compensate/fix it in any image from any digital camera.

Roger

[Emil Martinec]

My point was you don't know the magnitude of the fixed pattern noise. For example, by subtracting two different filters, you don't know if you have increase it by 2x or completely suppressed it. It seems to me that a better way is to model the low frequency background in single images and ratio it out, leaving the shot noise and fixed pattern noise.

Roger

I would agree with that. It's hardly worth the effort, but the way to go is probably to FT the image and throw out the lower frequencies, using the high frequency data to infer the local pixel variation. However, unless there is a pairwise correlation of adjacent columns I don't think there is going to be any special cancellation or reinforcement of values. A visual inspection of the sample I posted indicates that this is not an issue.

Personally, I don't see PRNU as an issue and have never had to compensate/fix it in any image from any digital camera.

Roger

I would agree with that. The only place where I would see it as an issue is where its patterned component might mislead and corrupt demosaicing. One might see it on the 7D as a corduroy effect in low ISO skies, but it's at such high SNR's that I doubt it.

The only reason I brought the subject up was your comment "The DXO curves you posted look fine for the Nikon, but are definitely not representative of the Canon. The Canon low iso (top curve) shows curvature over the entire range, while the Nikon is straight at the top 2/3rds or so. The only conclusion I can draw from that is that DXO does not appear to use actual raw data."

I think the DxO data is quite well explained at the low ISO top end by PRNU effects, which in their documentation they explicitly state is part of what their methodology tests for.

[Pao Dolina]

Gentlemen, bottomline given a choice of your Canon setup would you switch to a D3S or upgrade to a 1D4?