Luminous Landscape Forum
Equipment & Techniques => Cameras, Lenses and Shooting gear => Topic started by: ErikKaffehr on November 07, 2012, 01:40:16 am
-
Hi,
I have published a small article named: MF Digital, myths or facts?
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts
It is intended to drill down in some of the issues, hopefully it is unbiased.
Best regards
Erik
-
Erik,
Thanks for the link.
For the first point about the # of photons reaching the sensor...
Yes, a large sensor will collect more photons is exposed to the same light.
Now... the only source of illumination of the sensor is the rear element of the lens, correct?
So I don't think you can comment on photon noise without considering the sensor/camera/lens as a system.
Cheers,
Bernard
-
Had a brief look. I would argue that such a technically detailed article needs references. For example, you claim "Readout noise for CCDs used in MFD digital is about 12 electron charges". Why should this be the case? How can I see that?
Some things are rather confusing, e.g. "If we assume that we have a full frame sensor of 24x36 mm and compare it with a MF sensor of 24x48 mm size the later[sp?] one will have twice the area, so it will collect about the same number of photons". Equal intensity assumed, twice the area gives twice the number of photos.
-
I think you've got potentially a really good article, Eric. I didn't get past the 2nd page; however.
Why?
A few things.
First, there appears to be an error in the statement about sensor area in the 2nd graf of page 2. You say a 24x48 sensor will have twice the area of a 24x36 sensor. Is that correct? 24x36=864 and 24x48=1152. That's only a 33% increase in sensor area. If my math is wrong, please let me know. I think the statement ".... the later one will have twice the area, so it will collect about the same number of photons." is confusing. I know what you mean, but a lot of people may not.
Additionally, there are some, what I'll call, leaps of logic that someone who doesn't know about all this would find difficult to follow. I think even someone with some knowledge may find some of it difficult to follow.
I agree with Fips that the statement of the sensor having a read noise of 12 for a CCD needs support. You also state further down the page that read noise is 15. I think the SNR 12.5% needs some explanation. What does the 12.5% represent?
The statement "That noise is kept down when increasing ISO by applying preamplification before Analogue Digital Conversion." I think also needs clarity. People without a background in the science of electronics won't understand what this means.
Bernard, are you saying that there would be, or that there is, a significant difference in the transmission of light between different lenses all else being equal?
-
Bernard, are you saying that there would be, or that there is, a significant difference in the transmission of light between different lenses all else being equal?
Bob,
There may be, but that's not what I am saying. I am saying that the the number of photons reaching the sensor is not only a function of its size.
So I don't think we can say that MF sensors necessarily have more photons reaching them without considering the lens put in front of them.
Cheers,
Bernard
-
Right. And that was the point of my question. If the lens put in front of the sensor impacts the number of photons hitting the sensor doesn't that have to result from it transmitting a significantly different number of photons of light?
-
Right. And that was the point of my question. If the lens put in front of the sensor impacts the number of photons hitting the sensor doesn't that have to result from it transmitting a significantly different number of photons of light?
That's one part of it, but you of course have before that the size of the front element of the lens relative to its lenght, also defined as its aperture and the aperture being used, right?
Cheers,
Bernard
-
Thanks for all comments from everyone. Seems I have a lot of work to do, but that will make the article more useful.
All discussions in the article are based on the assumption that the FWC of the sensor is fully utilized. This may not be evident in the writing, sorry!
Best regards
Erik
That's one part of it, but you of course have before that the size of the front element of the lens relative to its lenght, also defined as its aperture and the aperture being used, right?
Cheers,
Bernard
-
That's one part of it, but you of course have before that the size of the front element of the lens relative to its length, also defined as its aperture and the aperture being used, right?
Cheers,
Bernard
True. But that's what I meant with my 'all else being equal' comment. Equal exposure in both cases. That should mean equal light being transmitted. Ignoring differences in aperture to get the same DOF on both formats, I'm just talking basic exposure.
-
True. But that's what I meant with my 'all else being equal' comment. Equal exposure in both cases. That should mean equal light being transmitted. Ignoring differences in aperture to get the same DOF on both formats, I'm just talking basic exposure.
If I remember right, the term we're looking for is "flux". Given equal flux density, a larger sensor will absorb more photons.
-
Well, I have no idea what 'flux density' is. Is it in any way related to the flux capacitor from "Back to the Future"?
But yes, I understand a larger sensor will capture more photons. That's not in question. Nor is it, I don't believe, related to the issue that Bernard raised and that I'm trying to get clarification on.
-
Yes, light transmission differs depending on systems, number of lenses, coatings, etc... and that plays a role in the endless discussions amateur astronomers have about their well characterized sensors (compared to digital cameras blackboxes in most cases). But I think that for photographic purposes, it is an acceptable approach to assume, in thoughts experiments, that all systems are exposed optimally, in other words exploit fully the linear zone of their sensors (even if they don't provide access to that level of raw data). Now I of course agree that if we go into sub-optimal low light exposures, long exposures etc... that assumption may not hold. Ultimately, all other things being equal, a bigger sensing area always wins. The main basis for our ongoing discussions is that all other things are never equal in this field/market and that by the time we have come to some kind of consensual evaluation, the target has moved as fast as a DSLR depreciates...
-
Erik, I think the weakest part of your great article is the section on color accuracy. There is so much that can influence the results, especially in the RAW processor. Every manufacturer imposes their idea of good color into a camera. I guess the best test would be how close could you get the cameras to a target by profiling and then see where the cameras differ from each other. And color accuracy has really two criteria, how accurate it is in absolute and relative terms. You can have very high absolute accuracy and really bad looking (unnatural) color.
-
The number of photos hitting a pixel of a certain area can be calculated by this:
n = (pi/4)(L*A/F^2(1+M)^2)*T optics * T atm. * t
Where
n = number of photons
L = spectral photon sterance
A = pixel area
F = f-number
M = magnification
T optics = Transmission of the lens
T atm. = Transmission of the atmosphere
t = exposure time
Basically, larger pixels gather more light given an equal f-number and exposure time.
-
It does not matter if a larger sensor "gathers" more light. It is irrelevant. What matters is the pixel size.
-
i don't know who your intended audience is but i reckon what you wrote is the cat's pyjamas. i read all of it from start to end and recommend others who do not have a phd yet but want to be able to speak at dinner parties as if they do should do the same. hopefully anybody who disagrees will make their own attempt to do as you have done that would help create a useful stream of readable literature on this subject instead of the usual tobacco industry marketing hype. of course what you are saying may be all smoke and mirrors but that is your right and much better than hiding behind a zillion pages of scientific gobblydook followed by 2 zillion pages of references that is for (yawn) academic journals not internet essays, right?
-
"It is possible that better results would been achieved by Capture One, but I'm pretty sure the comparison is pretty relevant regarding raw image data."
Based on knowing you here from the forum I assume it's a well researched, wel intended piece, written with minimal bias. But if the center of the discussion is centered on the "raw image data" then I lose interest pretty quickly.
Photographers show end-results, not raw pixels. If using a manufacturer's software (free to use with their backs) to process the images makes the end-result better (based on whatever criterium are important to the photographer) then that software should be considered an essential element of any analysis of the "myths or facts" of what makes the camera system sing.
Especially when a strong majority of Phase users use Capture One to process their Phase files (and many, maybe even most, use it to process their dSLR files as well).
-
There is a lot of good information here, but it is presented in a somewhat disjointed fashion, like you were just writing down good stuff as you thought of it. I think the article needs a second draft, editing the content to make it well organized and coherent. Maybe it would help to write an outline to define the structure of the article? It also needs a lot more detailed explanation if you want it to make sense to a non-expert audience.
Some notes on your "collecting more photons" section:
"If we assume that we have a full frame sensor of 24x36 mm and compare it with a MF sensor of 24x48 mm size the later one will have twice the area, so it will collect about the same number of photons." I have no idea what you mean here. Apart from the math error, don't you mean to say that the larger sensor will collect more photons? Or are you assuming something about relative pixel sizes and talking about photons per pixel?
Throughout the article, you never make a proper distinction between sensor size and pixel size. All of your dynamic range discussion is at the pixel level. Does the larger sensor have an advantage only because it provides larger pixels? (The answer is no, but I don't think I could conclude that from anything in your article.)
This is a technical article, so you should get the details right. For example, one of your conclusions is "A larger sensor will collect more photons and therefore have less shot noise." This is incorrect (as was shown in the earlier discussion). A correct statement would be "A larger pixel will collect more photons at a given gray level, providing a better signal/noise ratio".
From the discussion up to this point, I do not think you are entitled to replace "pixel" with "sensor" in the preceding statement, although it seems important to be able to do so. That reasoning comes next in your article, but only as a passing comment in an example.
Your discussion about pixels per printed image size comes late in the article but deserves more prominence. This is precisely why a large sensor has an advantage over a small sensor (irrespective of pixel size). It is not "software binning". It has a larger affect on the darks (where S/N is poor) than on the grays (where S/N is good).
-
Hi,
Please note that the first line of the article says: "Note: This is an article in progress".
There are some typing mistakes, like the one you have pointed out. I'll correct errors when they are found.
Regarding the number of photons collected, the only factor that really matters is the number of photons collected. Smaller pixels would collect fewer photons, but there would be more pixels. It matters very little if you collect 24 000 000 x 1000 photons or 6 000 000 x 4000 photons you still end up 24 000 000 000 000 photons. Would you print the image at 8x10" at 360 PPI you would end up with 2300 photons/pixel in both cases.
Once a print scale is fixed photons/pixel in the sensor is irrelevant.
There is some relevance to pixel size regarding DR but none at all with regard to shot noise.
Look at the included figures, the first one shows the effects of "sensor plus" on DR, there is a small improvement of DR above 400 ISO where Sensor+ is engaged (reducing resolution to 20 MP). The second figure shows that there is very little effect of "Sensor+" on tonal range, which is dominated by shot noise.
Best regards
Erik
There is a lot of good information here, but it is presented in a somewhat disjointed fashion, like you were just writing down good stuff as you thought of it. I think the article needs a second draft, editing the content to make it well organized and coherent. Maybe it would help to write an outline to define the structure of the article? It also needs a lot more detailed explanation if you want it to make sense to a non-expert audience.
Some notes on your "collecting more photons" section:
"If we assume that we have a full frame sensor of 24x36 mm and compare it with a MF sensor of 24x48 mm size the later one will have twice the area, so it will collect about the same number of photons." I have no idea what you mean here. Apart from the math error, don't you mean to say that the larger sensor will collect more photons? Or are you assuming something about relative pixel sizes and talking about photons per pixel?
Throughout the article, you never make a proper distinction between sensor size and pixel size. All of your dynamic range discussion is at the pixel level. Does the larger sensor have an advantage only because it provides larger pixels? (The answer is no, but I don't think I could conclude that from anything in your article.)
This is a technical article, so you should get the details right. For example, one of your conclusions is "A larger sensor will collect more photons and therefore have less shot noise." This is incorrect (as was shown in the earlier discussion). A correct statement would be "A larger pixel will collect more photons at a given gray level, providing a better signal/noise ratio".
From the discussion up to this point, I do not think you are entitled to replace "pixel" with "sensor" in the preceding statement, although it seems important to be able to do so. That reasoning comes next in your article, but only as a passing comment in an example.
Your discussion about pixels per printed image size comes late in the article but deserves more prominence. This is precisely why a large sensor has an advantage over a small sensor (irrespective of pixel size). It is not "software binning". It has a larger affect on the darks (where S/N is poor) than on the grays (where S/N is good).
-
But if the center of the discussion is centered on the "raw image data" then I lose interest pretty quickly.
give the business you are in somehow nobody is surprised
-
Hi,
Sorry it should be twice the photons.
Regarding the readout noise I have chosen 12 electrons/pixel because I wanted to use an optimistic value, I believe that 15-20 range is more probable for MFDB. Just a few lines below an example is given that is based on data from sensorgen, I will include a reference in the next revision.
Best regards
Erik
Had a brief look. I would argue that such a technically detailed article needs references. For example, you claim "Readout noise for CCDs used in MFD digital is about 12 electron charges". Why should this be the case? How can I see that?
Some things are rather confusing, e.g. "If we assume that we have a full frame sensor of 24x36 mm and compare it with a MF sensor of 24x48 mm size the later[sp?] one will have twice the area, so it will collect about the same number of photons". Equal intensity assumed, twice the area gives twice the number of photos.
-
Ultimately, all other things being equal, a bigger sensing area always wins.
but the hard fact is that technology for smaller sensors (dSLR/PS/cell phones) is nowadays consistently not the same as for bigger sensors in MF(DB)... so we can't consider all other things equal and operate just by sensor area alone (so while it wins in some, in certain areas it doesn't)
-
give the business you are in somehow nobody is surprised
I would imagine not. Given that my business is helping photographers pick cameras to make pictures. Not to help scientists diagnose quantum efficiency values.
Don't get me wrong. I think it's a great article. I'm just not in the target audience, nor do I think more than a few % of my customers would be. At least not as written/geared now.
-
Hi,
My take is that Doug's observation is a valid one. It is about the whole package. I have great respect for Doug and I think that he has made great contributions to these forums.
My intention with the article is to put things in some perspective. Now, we can have different perspectives, depending on experience. I just try to present some facts, keeping bias to a minimum.
Let us just take an example. There is something called thermal noise. It is my understanding that Phase One raw files contain info about ambient temperature, and perhaps also sensor temperature. Phase can use that information to selectively reduce noise in the darks. Unfair advantage to Phase? Probably! Do other vendors do similar things? Probably!
Best regards
Erik
I would imagine not. Given that my business is helping photographers pick cameras to make pictures. Not to help scientists diagnose quantum efficiency values.
Don't get me wrong. I think it's a great article. I'm just not in the target audience, nor do I think more than a few % of my customers would be. At least not as written/geared now.
-
Regarding the number of photons collected, the only factor that really matters is the number of photons collected. Smaller pixels would collect fewer photons, but there would be more pixels. It matters very little if you collect 24 000 000 x 1000 photons or 6 000 000 x 4000 photons you still end up 24 000 000 000 000 photons. Would you print the image at 8x10" at 360 PPI you would end up with 2300 photons/pixel in both cases.
So if you cut an image in half, you have half the number of photons? But so what. What is visible is still the same. (Lets ignore the fact that prints and files don't have photons.)
And don't you care about the well capacity and how many photons get in that well? This is what signal is after all. A pixel with a small signal is still a pixel with a small signal--what is around it does not change it into a pixel with more of a signal.
The pixel is a luminance/color data point. It has no spacial information beyond its position in the array. That luminance/color value is directly related to how many photon strikes it receives. The pixels around it are unrelated. S/N of the pixel is important.
-
Hi,
A pixel is just a number. By itself it can have no noise.
If you cut an image in half it will loose half of the photons. If you print the same size it will be 41 percent noisier. It is exactly the same as underexposing one stop.
Using a smaller image you also loose resolution.
Best regards
Erik
So if you cut an image in half, you have half the number of photons? But so what. What is visible is still the same. (Lets ignore the fact that prints and files don't have photons.)
And don't you care about the well capacity and how many photons get in that well? This is what signal is after all. A pixel with a small signal is still a pixel with a small signal--what is around it does not change it into a pixel with more of a signal.
The pixel is a luminance/color data point. It has no spacial information beyond its position in the array. That luminance/color value is directly related to how many photon strikes it receives. The pixels around it are unrelated. S/N of the pixel is important.
-
Hi Doug,
Thanks for the nice comment!
I would be glad for suggestions about gearing/writing, my intention is to be objective.
Best regards
Erik
I would imagine not. Given that my business is helping photographers pick cameras to make pictures. Not to help scientists diagnose quantum efficiency values.
Don't get me wrong. I think it's a great article. I'm just not in the target audience, nor do I think more than a few % of my customers would be. At least not as written/geared now.
-
Hi,
A pixel is just a number. By itself it can have no noise.
But it does have a signal. Are you saying signal is irrelevant? Distributing the exposure over the entire DR of the pixel does not matter?
If you cut an image in half it will loose half of the photons. If you print the same size it will be 41 percent noisier. It is exactly the same as underexposing one stop.
Now you are stretching. S/N ratio does not change. And since the signal does not change, neither does the exposure.
Using a smaller image you also lose resolution.
Now, you are changing the topic.
-
Hi,
What I have seen from the my work behind the article is that larger sensors would have smoother tones in light gray areas and would have advantage in edge contrast on fine detail.
Best regards
Erik
but the hard fact is that technology for smaller sensors (dSLR/PS/cell phones) is nowadays consistently not the same as for bigger sensors in MF(DB)... so we can't consider all other things equal and operate just by sensor area alone (so while it wins in some, in certain areas it doesn't)
-
Isn't it true that both the pixel well capacity and the size of the sensor matter? The pixel well capacity plays a larger part in the dynamic range. The size of the sensor plays a larger part in the total number of photons captured and thus the overall signal to noise ratio. So, Eric, in your example, while the two sensors would capture the same number of photons and could have the same S/N as a result, the sensor with the larger pixels should exhibit a better overall dynamic range. Correct?
Deeejjjaaa is also correct that not everything is equal between the different formats. As you point out, Eric, the noise characteristics of CCD and CMOS sensors are different so a direct comparison is somewhat difficult. Technology such as back-side illumination is also advantageous, allowing more light to be captured (a) in each pixel and (b) in total.
-
Hi,
Yes, a larger pixel would have a small advantage in DR. The reason that this is often not the case is that the sensor chips with the highest resolution often use a more advanced chip technology, leading to lower readout noise.
Back side illumination has many advantages, but I don't think it has been implemented in DSLRs yet.
Best regards
Erik
Isn't it true that both the pixel well capacity and the size of the sensor matter? The pixel well capacity plays a larger part in the dynamic range. The size of the sensor plays a larger part in the total number of photons captured and thus the overall signal to noise ratio. So, Eric, in your example, while the two sensors would capture the same number of photons and could have the same S/N as a result, the sensor with the larger pixels should exhibit a better overall dynamic range. Correct?
Deeejjjaaa is also correct that not everything is equal between the different formats. As you point out, Eric, the noise characteristics of CCD and CMOS sensors are different so a direct comparison is somewhat difficult. Technology such as back-side illumination is also advantageous, allowing more light to be captured (a) in each pixel and (b) in total.
-
Hi,
1) A pixel is just a number, like 1124. Does a number have noise?
2) I'm not stretching. If you crop the image it will be enlarged more. So each pixel in the print will "see" less photons, noise will increase.
Best regards
Erik
But it does have a signal. Are you saying signal is irrelevant? Distributing the exposure over the entire DR of the pixel does not matter?
Now you are stretching. S/N ratio does not change. And since the signal does not change, neither does the exposure.
Now, you are changing the topic.
-
Hi,
Yes, a larger pixel would have a small advantage in DR. The reason that this is often not the case is that the sensor chips with the highest resolution often use a more advanced chip technology, leading to lower readout noise.
And I think this is the point deeejjjaaa was making.
Back side illumination has many advantages, but I don't think it has been implemented in DSLRs yet.
Best regards
Erik
Not sure about DSLRs. Certainly P&S cameras and cell phones. Maybe some mirrorless type cameras, not sure. But this gets back to the point about how different technologies can impact the final analysis.
[/quote]
-
Regarding the number of photons collected, the only factor that really matters is the number of photons collected. Smaller pixels would collect fewer photons, but there would be more pixels. It matters very little if you collect 24 000 000 x 1000 photons or 6 000 000 x 4000 photons you still end up 24 000 000 000 000 photons. Would you print the image at 8x10" at 360 PPI you would end up with 2300 photons/pixel in both cases. Once a print scale is fixed photons/pixel in the sensor is irrelevant.
I agree completely. My point was that your article does not make this clear, since the discussion is almost all about photons per pixel. Adding the content above to your article in progress would be a great improvement.
-
Hi Eric,
Thanks a lot. Suggestions like yours are most helpful ;-)
Best regards
Erik
I agree completely. My point was that your article does not make this clear, since the discussion is almost all about photons per pixel. Adding the content above to your article in progress would be a great improvement.
-
Hi!
I just added a comment from Doug Peterson from Digital Transitions here: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts?start=11
Doug is a long time member of LuLa with an endless number of good contributions.
Best regards
Erik
Hi,
I have published a small article named: MF Digital, myths or facts?
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts
It is intended to drill down in some of the issues, hopefully it is unbiased.
Best regards
Erik
-
Hi,
I'd suggest that there are some advantages to back illuminated CMOS like better fill factor and less lens cast effects, but a change of technology would have little effect on the analysis. It would foremost give better low light performance and that is not included in the article. I may add some info about the issues.
Thanks very much for your input!
Best regards
Erik
And I think this is the point deeejjjaaa was making.
Not sure about DSLRs. Certainly P&S cameras and cell phones. Maybe some mirrorless type cameras, not sure. But this gets back to the point about how different technologies can impact the final analysis.
-
And don't you care about the well capacity and how many photons get in that well? This is what signal is after all. A pixel with a small signal is still a pixel with a small signal--what is around it does not change it into a pixel with more of a signal.
The pixel is a luminance/color data point. It has no spacial information beyond its position in the array. That luminance/color value is directly related to how many photon strikes it receives. The pixels around it are unrelated. S/N of the pixel is important.
A pixel is not just a luminance data point at a particular location, it also represents a certain subject size. One way to think of this is by taking the angular coverage of the lens and dividing by the number of pixels covered. Another way to think of this is by considering how many pixels are occupied by a physical object in the scene. By representing a physical object (say a patch of uniform gray sky) with more pixels, I get inherently better signal/noise in the final printed image. It is not just noise per pixel that matters.
Under the assumptions that S/N per pixel is dominated by the number of photons collected, and that the number of photons collected per pixel is proportional to the pixel area, a sensor of a given size will have the same overall noise performance whether I divide the sensor area into small pixels or large pixels. The smaller pixels will have worse S/N per pixel, but in the final image that will be precisely compensated for by the increased number of pixels.
-
1) A pixel is just a number, like 1124. Does a number have noise?
When your pixel states 1124, it simply means that on a certain sensor you can expect the actual incoming data that made that pixel value was (for example) between 1114 and 1134. The neighbouring pixel wich could be measuring the same data (say a uniform background) could have 1113 +/-10 and another one, slightly less sensitive 983 +/-11. You'd then see a "noisy" band of three pixels in place of the uniform one you were hoping for.
If you want to state it in another way, there is a margin of incertitude on that single value, and that uncertain part of the signal is the noise in the signal.
2) I'm not stretching. If you crop the image it will be enlarged more. So each pixel in the print will "see" less photons, noise will increase.
If, in the simplest method, the pixel is doubled. That doubling doesn't change the SNR. You are simply increasing the area that represents the 1124 (+/-10) measurement. Neither changing its value, nor changing the error margin that occurred when it was captured.
-
area, a sensor of a given size will have the same overall noise performance whether I divide the sensor area into small pixels or large pixels. The smaller pixels will have worse S/N per pixel, but in the final image that will be precisely compensated for by the increased number of pixels.
In practice, this works roughly for digital cameras (as demonstrated in Emil Martinec's paper).
Now, consider the following cameras that happen to have 1 unit of read noise.
Camera 1 has a sensel that collects 16 units of lights. When read, it will be 15 - 16 - 17
Camera 2 has 4 sensels that collect 4 units of lights each. When read, you'll have 3-4-5 / 3-4-5 / 3-4-5 / 3-4-5 (worst case 12 - 20)
So you see that this is not "precisely compensated" although the distribution of your signal will indeed be centered on the same 15-16-17.
It doesn't matter too much with photography because you are almost always using a decent exposure and working with a significantly larger number of units than the ones above. So you clould simply say that you'll ignore the issue because it is negligible (do the math with 60000 units and 4 times 15000 units and a 10 units read noise for example) but redefining noise or demonstrating an equality that doesn't exist will not satsify everyone.
The noise (uncertainty on the captured signal) doesn't change after the capture if you can store the data reliably.
One last thing: the "scientific" definition of noise and the "photographic" perception of noise are two different things: shooting a deep dark sky background should be noisy simply because the sky background is not uniform (there's also the variability rates or light unit arrivals, Poisson, but le't's not get into that) but a photographer will find the totally uniform black background produced by the Nikon blackbox less noisy.
(fixed typos)
-
... So you see that this is not "precisely compensated" ...
Which is why I was careful to preface my comment with the disclaimer "Under the assumptions that S/N per pixel is dominated by the number of photons collected ..." :)
As you point out, when readout noise is significant, there is a penalty for more numerous smaller pixels, which have a noise component which does not scale with pixel size or collected photons.
-
Hi,
If you check this posting: http://www.luminous-landscape.com/forum/index.php?topic=72167.msg572836#msg572836
You can see that quadrupling the pixel area has a small effect on DR when normalized to print size but a very small effect on tonal range.
For me this is a good indication that EricV is right.
Best regards
Erik
Which is why I was careful to preface my comment with the disclaimer "Under the assumptions that S/N per pixel is dominated by the number of photons collected ..." :)
As you point out, when readout noise is significant, there is a penalty for more numerous smaller pixels, which have a noise component which does not scale with pixel size or collected photons.
-
Hi,
Yes, I agree, but it is pretty much what I say in the article. There is a comparison of two exposures of a color checker which I presume were taken under identical conditions.
My understanding is that color is either accurate or pleasant. Both of the images I tested were oversaturated (technically speaking) and I reduced saturation on both to get close to correct saturation. The article says: "The measured data above actually indicates that the D800E is better in reproducing a color checker card under a given set of conditions. The main difference between the Hassy and the D800E was that the Hassy image processed in LR4.2 was significantly oversaturated. When processing in LR4.2 I pulled back 13 units of saturateion on the Hassy and 4 units on the Nikon. Delta E is about half on the Nikon."
I tried to profile the CC-shots I used but they were both slightly overexposed.
Best regards
Erik
Erik, I think the weakest part of your great article is the section on color accuracy. There is so much that can influence the results, especially in the RAW processor. Every manufacturer imposes their idea of good color into a camera. I guess the best test would be how close could you get the cameras to a target by profiling and then see where the cameras differ from each other. And color accuracy has really two criteria, how accurate it is in absolute and relative terms. You can have very high absolute accuracy and really bad looking (unnatural) color.
-
Which is why I was careful to preface my comment with the disclaimer "Under the assumptions that S/N per pixel is dominated by the number of photons collected ..." :)
As you point out, when readout noise is significant, there is a penalty for more numerous smaller pixels, which have a noise component which does not scale with pixel size or collected photons.
I think we agree, which is why I said that it doesn't matter much for photographic applications with decent exposure. It still does matter slightly for the dark areas of a properly exposed picture, which is why Nikon is essentially castrating them for a signal processing point of view. It did of course matter more when small pixels (say 15000 / 25000 FWC) where suffering from higher read noise (say 15)...
What tickles me a bit is when instead of saying "for a whole lot of complex reasons it ends up not mattering much in practice and the result is roughly equivalent" one says "this is precisely so as demonstrated there".
Very minor issue, I concede, and in the context of photographic education, your approach probably beats mine.
-
1) A pixel is just a number, like 1124. Does a number have noise?
If that number is used to represent some real number between 0 and 2000 (e.g. 1124.42), then I would say that it is a noisy measurement, yes.
-h
-
Under the assumptions that S/N per pixel is dominated by the number of photons collected, and that the number of photons collected per pixel is proportional to the pixel area, a sensor of a given size will have the same overall noise performance whether I divide the sensor area into small pixels or large pixels. The smaller pixels will have worse S/N per pixel, but in the final image that will be precisely compensated for by the increased number of pixels.
This is not exactly correct, since if one bins 4 pixels into one pixel post capture via software, the binned superpixel will have 4 read noise contributions whereas a larger pixel with 4x the area would have only one read noise contribution. Software binning is the mechanism underlying the DXO screen vs pixel data. Hardware binning is widely used with monochrome scientific CCDs (see here (http://www.photometrics.com/resources/learningzone/binning.php)), but the process is considerably more complex for Bayer array sensors and as far as I know, hardware binning with Bayer sensors is only available with the Phase One sensor plus technology (see here (http://www.phaseone.com/en/Camera-Systems/IQ-Series/IQ-Tutorials.aspx). Click on the P+ tutorial).
While a large sensor does collect more photoelectrons, one should remember that the SNR contribution from shot noise increases as the square root of the number of photons collected. Doubling the sensor area (as in going from an APS sized sensor to a full frame 35 mm sensor) will improve the SNR by a factor of only 1.4. Newer technology CMOS sensors (such as in the Nikon D7000) can compete quite well with older full frame sensors. The same considerations apply to MFDBs. As Erik has pointed out, the MFDBs are hampered by their high read noise which limits their dynamic range. However, their SNR in the midtones (where shot noise does not contribute significantly to the SNR) is quite good.
For ultimate image quality few well informed observers would deny that MFDBs are the way to go, but the price to performance ratio is quite steep.
Regards,
Bill
-
Hi,
I guess that my findings agree pretty well with Bill's conclusion. MFDB has a small advantage regarding shot noise in highlights and midtones. I would suggest that this may be hard to illustrate with images, because all the modern sensors are pretty good in this area.
I would expect that MFDBs would respond better to sharpening compared to DSLRs because I would expect them to have less shot noise. MFDBs are normally not OLP-filtered and they normally don't have microlenses, which may also reduce the need of sharpening.
You really need to look at the whole package. I'm pretty sure that MFDBs have an advantage in the resolution/MTF/microcontrast area. On the other hand I suspect that the DR advantage of MF is by and large a myth. Color reproduction and midtone tonality, I don't know.
Best regards
Erik
This is not exactly correct, since if one bins 4 pixels into one pixel post capture via software, the binned superpixel will have 4 read noise contributions whereas a larger pixel with 4x the area would have only one read noise contribution. Software binning is the mechanism underlying the DXO screen vs pixel data. Hardware binning is widely used with monochrome scientific CCDs (see here (http://www.photometrics.com/resources/learningzone/binning.php)), but the process is considerably more complex for Bayer array sensors and as far as I know, hardware binning with Bayer sensors is only available with the Phase One sensor plus technology (see here (http://www.phaseone.com/en/Camera-Systems/IQ-Series/IQ-Tutorials.aspx). Click on the P+ tutorial).
While a large sensor does collect more photoelectrons, one should remember that the SNR contribution from shot noise increases as the square root of the number of photons collected. Doubling the sensor area (as in going from an APS sized sensor to a full frame 35 mm sensor) will improve the SNR by a factor of only 1.4. Newer technology CMOS sensors (such as in the Nikon D7000) can compete quite well with older full frame sensors. The same considerations apply to MFDBs. As Erik has pointed out, the MFDBs are hampered by their high read noise which limits their dynamic range. However, their SNR in the midtones (where shot noise does not contribute significantly to the SNR) is quite good.
For ultimate image quality few well informed observers would deny that MFDBs are the way to go, but the price to performance ratio is quite steep.
Regards,
Bill
-
Hi,
My point is that a pixel in isolation is pretty meaningless. Noise arises when there are several dozens or hundred pixels.
On pixel with 1124 electron charges noise would be 33.5 electron charges, so 65% of the pixels would hold between 1088 and 1158 electron charges (if I recall correctly).
Best regards
Erik
If that number is used to represent some real number between 0 and 2000 (e.g. 1124.42), then I would say that it is a noisy measurement, yes.
-h
-
Had a brief look. I would argue that such a technically detailed article needs references. For example, you claim "Readout noise for CCDs used in MFD digital is about 12 electron charges". Why should this be the case? How can I see that?
Some things are rather confusing, e.g. "If we assume that we have a full frame sensor of 24x36 mm and compare it with a MF sensor of 24x48 mm size the later[sp?] one will have twice the area, so it will collect about the same number of photons". Equal intensity assumed, twice the area gives twice the number of photos.
I thought that the IQ180 sensor size was 40.4x53.7mm, not 24x48mm ?
And, please Erik, how do you calculate SNR 12.5 ?!
Thierry
-
My point is that a pixel in isolation is pretty meaningless. Noise arises when there are several dozens or hundred pixels.
That pixel (sensel) in isolation is however how all sensors' DR is characterized. FWC, read noise.... You can have a one sensel sensor, a million sensels sensor...
In your definition, where do you put the limit when noise arises? 24 pixels? 36? 48? 480?
If there is a certain amount of noise at 500 pixels, does it go up or down at 1000 pixels?
Assuming by that definition that there is x amount of noise with y pixels, what's the noise for one pixel x/y or x.y?
If there is no noise (zero noise) in the one pixel case how come 1000 of them have some amount of noise at all?
-
Hi!
Please check the enclosed pixel and tell me the noise level.
If we make an image of an evenly illuminated surface there will be a statistical variation on the pixels. If we presume the data numbers correspond to electron charges we would now that 64% of the pixels would be within 1090 and 1159 electron charges if the mean was 1124. But a single pixel doesn't have noise.
Best regards
Erik
That pixel (sensel) in isolation is however how all sensors' DR is characterized. FWC, read noise.... You can have a one sensel sensor, a million sensels sensor...
In your definition, where do you put the limit when noise arises? 24 pixels? 36? 48? 480?
If there is a certain amount of noise at 500 pixels, does it go up or down at 1000 pixels?
Assuming by that definition that there is x amount of noise with y pixels, what's the noise for one pixel x/y or x.y?
If there is no noise (zero noise) in the one pixel case how come 1000 of them have some amount of noise at all?
-
Hi,
Added a sample images of Phase One IQ 180 processed in LR42 and Capture 1 compared with Nikon D800.
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts?start=3
Best regards
Erik
Hi,
I have published a small article named: MF Digital, myths or facts?
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/71-mf-digital-myths-or-facts
It is intended to drill down in some of the issues, hopefully it is unbiased.
Best regards
Erik
-
If we make an image of an evenly illuminated surface there will be a statistical variation on the pixels. If we presume the data numbers correspond to electron charges we would now that 64% of the pixels would be within 1090 and 1159 electron charges if the mean was 1124. But a single pixel doesn't have noise.
Single pixel noise is a useful concept and does make sense. For a group of similar pixels with similar illumination, we agree that noise causes statistical variation in the response across the different pixels. Similarly for a single pixel, I think we can agree that noise will cause statistical variation in the response of that pixel over time or over repeated measurements. It is not a great conceptual leap to say that a single pixel has a true value and a measurement error. The true value is the average light intensity, while the error is the statistical fluctuation expected for that light intensity. For a single image the best estimate of the true value is the measured value, but many noise reduction techniques rely on capturing multiple images to provide an improved estimate of the true value.
Instead of saying simply "this pixel has a measured value of 1124" it would be more informative to say "this pixel has a measured value of 1124 and a statistical uncertainty of 33". In scientific publications, when measurement results are reported, it is rather common to see statements like "pixel value = 1124 +- 33". If there are multiple sources of error, it is even better to say something like "pixel value = 1124 +- 33 (statistical) +- 12 (readout noise)".
-
Hi,
Yes, I absolutely agree with that reasoning. On the other hand you are still sampling photons so you essentially says the pixel varies over time. It is pretty similar to the statistical variation over a simultaneous sampling over a number of pixels. You still need several samples to see a variation.
My point, mostly, is that we never do anything useful with a single pixel. We always use a large number of pixels and there will be a statistical variation.
You are right about the readout noise. It would be significant at 1124 electron charges on older sensors. On the latest CMOS sensors readout noise seems to be around 3 electron charges. I also think that you would add noise in quadrature. So shot noise which would be around 34 charges would dominate.
Best regards
Erik
Single pixel noise is a useful concept and does make sense. For a group of similar pixels with similar illumination, we agree that noise causes statistical variation in the response across the different pixels. Similarly for a single pixel, I think we can agree that noise will cause statistical variation in the response of that pixel over time or over repeated measurements. It is not a great conceptual leap to say that a single pixel has a true value and a measurement error. The true value is the average light intensity, while the error is the statistical fluctuation expected for that light intensity. For a single image the best estimate of the true value is the measured value, but many noise reduction techniques rely on capturing multiple images to provide an improved estimate of the true value.
Instead of saying simply "this pixel has a measured value of 1124" it would be more informative to say "this pixel has a measured value of 1124 and a statistical uncertainty of 33". In scientific publications, when measurement results are reported, it is rather common to see statements like "pixel value = 1124 +- 33". If there are multiple sources of error, it is even better to say something like "pixel value = 1124 +- 33 (statistical) +- 12 (readout noise)".
-
Hi!
Please check the enclosed pixel and tell me the noise level.
Hi Erik,
Shoot it several times, or read it several times after resetting it, and its noise can be calculated ...
A noiseless sensel will produce the same output every time (ain't gonna happen).
Cheers,
Bart
-
Hi Bart,
Same as EricV says and I agree. My point is more like that a pixel is pretty meaningless without a context. With multiple exposures you add a temporal context, but I still think that very few of us would enjoy a single pixel movie, although the pixel would have both shot noise and readout noise. A couple of millions of those pixels on the other hand give a nice image.
It would be feasible to build a sensor that has binary pixels either black or white. If there was enough of those pixels the sensor would form a good image. As far as I know such sensor designs have been proposed.
Best regards
Erik
Hi Erik,
Shoot it several times, or read it several times after resetting it, and its noise can be calculated ...
A noiseless sensel will produce the same output avery time (ain't gonna happen).
Cheers,
Bart
-
It would be feasible to build a sensor that has binary pixels either black or white. If there was enough of those pixels the sensor would form a good image. As far as I know such sensor designs have been proposed.
And realized, one often cited paper is
http://www.ece.rice.edu/~duarte/images/csCamera-SPMag-web.pdf
you might be surprised by the results.
-
Hi,
Sorry for taking time responding.
SNR 12.5 is meant as SNR 3 stops below saturation, 2^3 = 8. If we divide 100% with eight we get 12.5%.
The 24x48 sensor should be 36x48 (two 24x36 sensors side by side) and this was given as an example of doubling sensor size. The example was absolutely hyptothetical.
The real calculation is based on P65+ as data for that sensor is available from sensorgen.
Best regards
Erik
I thought that the IQ180 sensor size was 40.4x53.7mm, not 24x48mm ?
And, please Erik, how do you calculate SNR 12.5 ?!
Thierry
-
Totally agree with you Bernard. The amount of light reaching the sensor will be a function of the gathering power of the lens, the internal transmission losses and the percentage of the image circle that actually falls on the sensor. Perhaps an interesting test would be to compare various 35mm lenses with various MFD lenses in a test rig on both types of sensors so that these variables can be eliminated.
I shoot with both MFD (H4D-60) and 35mm (D800E). Both are fine instruments and very often I could use either camera for a job. There are situations however where the ease of use and portability of the Nikon make it my tool of choice and situations where the Hasselbald is my preferred option - usually in the studio. They are both very very good.
However from a subjective point of view I like what the combination of Hasselblad back and lenses and their Phocus software producing in a large format print.
-
Hi,
Sorry for taking time responding.
SNR 12.5 is meant as SNR 3 stops below saturation, 2^3 = 8. If we divide 100% with eight we get 12.5%.
The 24x48 sensor should be 36x48 (two 24x36 sensors side by side) and this was given as an example of doubling sensor size. The example was absolutely hyptothetical.
The real calculation is based on P65+ as data for that sensor is available from sensorgen.
Best regards
Erik
Thanks Erik !
Thierry
-
Hi,
Actually that does not matter at all as far as you expose consistently, especially not if you expose to the right, thus fully utilizing the sensor.
Best regards
Erik
Totally agree with you Bernard. The amount of light reaching the sensor will be a function of the gathering power of the lens, the internal transmission losses and the percentage of the image circle that actually falls on the sensor. Perhaps an interesting test would be to compare various 35mm lenses with various MFD lenses in a test rig on both types of sensors so that these variables can be eliminated.
I shoot with both MFD (H4D-60) and 35mm (D800E). Both are fine instruments and very often I could use either camera for a job. There are situations however where the ease of use and portability of the Nikon make it my tool of choice and situations where the Hasselbald is my preferred option - usually in the studio. They are both very very good.
However from a subjective point of view I like what the combination of Hasselblad back and lenses and their Phocus software producing in a large format print.
-
Well, I have no idea what 'flux density' is. Is it in any way related to the flux capacitor from "Back to the Future"?
Luminous flux = Lumens (http://en.wikipedia.org/wiki/Luminous_flux)
Exposure (http://en.wikipedia.org/wiki/Exposure_(photography)) = Lumens * integration time (lux seconds)
-
It would be feasible to build a sensor that has binary pixels either black or white. If there was enough of those pixels the sensor would form a good image. As far as I know such sensor designs have been proposed.
Such sensors exist: they are called "film", where each sensel is a cluster of silver halide crystals that is either "exposed" or "not exposed" as far as the subsequent processing is concerned. All shades of gray seen in a traditional monochrome negative are dithering of a mixture of pure black at exposed "sensels" and pure white elsewhere. By any practically relevant definition, the resulting dynamic range is vastly greater than any "per sensel" value, and cannot determined at all from per sensel measurements. Dithering and blurring of the signal from multiple sensels each covering an angular part of the viewed image too small for the viewer's eye to resolve must be taken into acccount.
By the way, thanks for your effort on this.
Also, the Kodak/Truesense 50MP CCD has read noise of 12.5 e-, about the best for any of that brand of sensors, so your value is reasonable for CCDs as used in DMF. One source of such data is
http://www.truesenseimaging.com/products/full-frame-ccd
and the product summaries linked to in there.
P. S. TeledyneDalsa just released details of a 60MP sensor which is presumably the same as or similar to the one that was originally exclusive to Phase One, only now offered openly to all customers:
http://www.teledynedalsa.com/public/sensors/datasheets/FTF9168C_datasheet_20120306.pdf
This reports signal and noise specs in a different way, but you can convert back from well capacity (50,000e-) and DR (73 dB) to SNR in electrons, and I get just over 11e-.
-
Thanks for the good comments!
Erik
Such sensors exist: they are called "film", where each sensel is a cluster of silver halide crystals that is either "exposed" or "not exposed" as far as the subsequent processing is concerned. All shades of gray seen in a traditional monochrome negative are dithering of a mixture of pure black at exposed "sensels" and pure white elsewhere. By any practically relevant definition, the resulting dynamic range is vastly greater than any "per sensel" value, and cannot determined at all from per sensel measurements. Dithering and blurring of the signal from multiple sensels each covering an angular part of the viewed image too small for the viewer's eye to resolve must be taken into acccount.
By the way, thanks for your effort on this.
Also, the Kodak/Truesense 50MP CCD has read noise of 12.5 e-, about the best for any of that brand of sensors, so your value is reasonable for CCDs as used in DMF. One source of such data is
http://www.truesenseimaging.com/products/full-frame-ccd
and the product summaries linked to in there.
P. S. TeledyneDalsa just released details of a 60MP sensor which is presumably the same as or similar to the one that was originally exclusive to Phase One, only now offered openly to all customers:
http://www.teledynedalsa.com/public/sensors/datasheets/FTF9168C_datasheet_20120306.pdf
This reports signal and noise specs in a different way, but you can convert back from well capacity (50,000e-) and DR (73 dB) to SNR in electrons, and I get just over 11e-.
-
Such sensors exist: they are called "film", where each sensel is a cluster of silver halide crystals
And the million dollar question obviously is: how many of those clumps fit in the space of a modern sensel at equivalent ISO?
-
And the million dollar question obviously is: how many of those clumps fit in the space of a modern sensel at equivalent ISO?
How many of those clumps can be fit into a film size that is economically viable to manufacture and shoot images with, compared to many sensels can be fit into a digital sensor that is economically viable to manufacture and shoot images with?
-h
-
And the million dollar question obviously is: how many of those clumps fit in the space of a modern sensel at equivalent ISO?
I am fairly sure that there are many billions of them in a 36x24mm frame of a typical fine-grained monochrome film. One hint is the very high extinction resolution of such films, indicating that the chemical sensels are far smaller than the photosites of any current DSLR sensor, though of course at that resolution limit, dynamic range and measurement of luminosity levels is very poor.
Eric Fossum, originator of the modern active pixel CMOS sensor, has proposed sensors based on one-bit electonic sensels, which he calls "jots". Some of his writings at his site http://ericfossum.com/ are
http://ericfossum.com/Presentations/2012%20March%20QIS%20London.pdf
where he suggests that one bilion to 100 billion one-bit sensels would be needed, and
http://ericfossum.com/Publications/Papers/Gigapixel%20Digital%20Film%20Sensor%20Proposal.pdf
He also speculates about repeated reading of sensels during an exposure.
-
Hi,
There are two problems with silver halide crystals. One is that quantum efficiency is low (about 2% if I recall) the other is that it takes a lot of crystals to replace a single pixel. A single pixel can take detect 30000-60000 photons. It would take 30000 crystals to detect 30000 photons.
Best regards
Erik
I am fairly sure that there are many billions of them in a 36x24mm frame of a typical fine-grained monochrome film. One hint is the very high extinction resolution of such films, indicating that the chemical sensels are far smaller than the photosites of any current DSLR sensor, though of course at that resolution limit, dynamic range and measurement of luminosity levels is very poor.
Eric Fossum, originator of the modern active pixel CMOS sensor, has proposed sensors based on one-bit electonic sensels, which he calls "jots". Some of his writings at his site http://ericfossum.com/ are
http://ericfossum.com/Presentations/2012%20March%20QIS%20London.pdf
where he suggests that one bilion to 100 billion one-bit sensels would be needed, and
http://ericfossum.com/Publications/Papers/Gigapixel%20Digital%20Film%20Sensor%20Proposal.pdf
He also speculates about repeated reading of sensels during an exposure.