How small can or should a pixel be

[ErikKaffehr]

Hi Marc,

Just a few comments:

1) Each pixel contains essentially a photo sensitive area and about 6 transistors. The size of the transistors depends on the design rules used. So if you decrease pixel size without using narrower design rules, the area sensor vs. electronics will shrink. The sensitive area also works as a capacitor, reducing the size of the capacitor will reduce the "full well capacity" of the sensor. The "full well capacity" is the determining factor for the high end of DR. Optimal pixel size varies with design rules.

This article explains it well: http://isl.stanford.edu/~abbas/group/papers_and_pub/pixelsize.pdf

2) I agree on the issue of avoiding the AA-filter, but with shrinking pixel size the thickness of the AA-filter will also shrink. I'd also say that the effects of AA-filtering are overstated as it is obvious that it is diffraction and not AA-filtering that limits lens sharpness. Once we stop down beyond optimum aperture resolution drops.

3) Binning in hardware makes the pixels larger. It is good for increasing DR, specially at high ISO, but doesn't counteract aliasing.

Best regards
Erik

Or how about a FF 120mp sensor that is binned to 30mp! no AA filter (the lens resolution becomes the filter) less artifacts etc
Marc

[Sareesh Sudhakaran]

Hi Erik

I was trying to read up on this last week and my understanding is:

1. Ideally a pixel should be as small as possible. The smaller it is, the better it can resolve.
2. However, the bigger the pixel is, the greater the dynamic range due to its light-collecting capacity.

Therefore, usually it's a trade-off. Am I right, or have I misunderstood completely?

[ErikKaffehr]

Hi,

1) Yes, smaller pixels will resolve more details, but there are limitations. The main limitation is diffraction, once a lens a is diffraction limited we get diminished returns from shrinking pixel size.

2) The DR of the pixel is reduced with it's size, but we get more pixels and that compensates a lot. LuLa has an excellent article on the issue here: http://www.luminous-landscape.com/essays/dxomark_sensor_for_benchmarking_cameras.shtml

Combining pixels in software doesn't reduce readout noise the same way as binning them in hardware.

Here is a demonstration of diffraction: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=1 left column shows effects of diffraction.

Best regards
Erik

Hi Erik

I was trying to read up on this last week and my understanding is:

1. Ideally a pixel should be as small as possible. The smaller it is, the better it can resolve.
2. However, the bigger the pixel is, the greater the dynamic range due to its light-collecting capacity.

Therefore, usually it's a trade-off. Am I right, or have I misunderstood completely?

[marcmccalmont]

Erik
A few months ago Canon announced the development of a 120mpx APSc sensor But like our 10 mpx point and shoots the real resolution is diffraction limited and resolution limited to much larger than those little point and shoot pixels right? But like on the IQ's use hardware binning and let the diffraction limit and resolution limit of the lens be the AA filter. in other words if the lens cannot resolve a single photosite but can resolve 4 photosites don't we have a simple solution? sharper pixels, better DR, no AA filter, simpler de-mosaicing (each group of 4 pixels has an RGB value of its own).
Marc

PS I vote for 2.5 micron photosites binned in fours! And let the sensor area determine the gross resolution

[Mark D Segal]

Erik, thanks for those references.

[uaiomex]

Please let me see if I understand. Smaller pixels create more detail than bigger pixels. The smaller the pixel the sooner it will fall into diffraction obliterating the extra resolution. The smaller the aperture, the more diffraction at any pixel size. Ok these.
As I understand, the detail obtained from bigger pixels will never be more than the detail obtained from smaller pixels no matter how much diffraction they suffer. Right? Or.. it reverses? I don't think so.
Thanks in advance
Eduardo

Hi,

1) Yes, smaller pixels will resolve more details, but there are limitations. The main limitation is diffraction, once a lens a is diffraction limited we get diminished returns from shrinking pixel size.

2) The DR of the pixel is reduced with it's size, but we get more pixels and that compensates a lot. LuLa has an excellent article on the issue here: http://www.luminous-landscape.com/essays/dxomark_sensor_for_benchmarking_cameras.shtml

Combining pixels in software doesn't reduce readout noise the same way as binning them in hardware.

Here is a demonstration of diffraction: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=1 left column shows effects of diffraction.

Best regards
Erik

[Sareesh Sudhakaran]

Thanks for the links, Erik. Appreciate it.

[ErikKaffehr]

Hi,

Yes, your observations are correct.

With smaller pixels diffraction effects will be more obvious. You always seem to loose some but with smaller pixels you loose more when stopping down. The image below has been measured on a 6MP DSLR and on one with 10 MP. The 10 MP camera always gives more detail, but both start loosing out at the same time.



Some part of what lost with diffraction can be regained with correct sharpening, the enclosed image should illustrate this. That image is taken from this article:
http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures

Best regards
Erik

Please let me see if I understand. Smaller pixels create more detail than bigger pixels. The smaller the pixel the sooner it will fall into diffraction obliterating the extra resolution. The smaller the aperture, the more diffraction at any pixel size. Ok these.
As I understand, the detail obtained from bigger pixels will never be more than the detail obtained from smaller pixels no matter how much diffraction they suffer. Right? Or.. it reverses? I don't think so.
Thanks in advance
Eduardo

[uaiomex]

Hi,
Erik: It's good to assure this "assumed" truth.
Muy agradecido
Eduardo

Yes, your observations are correct.
Best regards
Erik

[torger]

A dream sensor for me would be full-frame 40-48 pixel for high res still life / studio, binned to 4:1 10-12 megapixel for hand-held and high ISO.

Today it seems like small pixel sensors perform well up to ISO3200, but you still need large pixel sensors (for example D3s) to get decent quality in the ISO6400 - ISO25600 range. It does not seem to be possible to combine the best ultra high ISO performance with small pixel sizes yet.

Having to choose, I prefer high res over high ISO for my uses. However, the ultra-high range can be very useful for journalism/wildlife/sports/indoor events and is also useful for me at times, so if hardware binning could improve performance from ISO6400 and up that would yield a great all-around sensor. If I could get an ISO12800 with decent quality (similar to what D3s does today) and have that incorporated in the auto-ISO mode then I think I would very rarely feel limited by ISO range.

[bjanes]

Hi Marc,

Just a few comments:

1) Each pixel contains essentially a photo sensitive area and about 6 transistors. The size of the transistors depends on the design rules used. So if you decrease pixel size without using narrower design rules, the area sensor vs. electronics will shrink. The sensitive area also works as a capacitor, reducing the size of the capacitor will reduce the "full well capacity" of the sensor. The "full well capacity" is the determining factor for the high end of DR. Optimal pixel size varies with design rules.

This article explains it well: http://isl.stanford.edu/~abbas/group/papers_and_pub/pixelsize.pdf

Eric,

The Stanford article you referenced is excellent. Their calculations were for a CMOS sensor using 0.35 micron technology, giving a fill factor of 30%. The optimum pixel size was 6.5 microns. If one were to use a smaller CMOS scaling one could improve the fill factor and the pixel size could be smaller. For a given fill factor, the use of microlenses can improve light sensitivity by focusing all the light of the pixel area on the active pixel site. However, the full well might not improve, since charge density would be the limiting factor.

CCDs have no electronics on the pixel, so their fill factor is larger.

With improved sensors with lower read noise, performance of smaller pixels at base ISO is often fully satisfactory.

Regards,

Bill

[theguywitha645d]

The amount of detail is not a simple function of pixel pitch. Which sensor would produce more detail, a 30MP 24x36mm sensor or a 30MP 48x72mm sensor? And when does detail become meaningless--if your viewer cannot perceive the detail or the output device unable to render the detail, what is the point of the detail?

[ErikKaffehr]

Hi,

Both sensors will produce the same amount of detail, but the 48x72 mm sensor will have better microcontrast. The 48x72 mm sensor will also produce a lot of fake detail unless it has a very agressive OLP filter or you stop it down to f/22-f/32. (The Leica S2 seems to cease aliasing when stopped down past f/11, your hypothetical 48x72 sensor would have almost the double pixel size so it would need stopping down two stops more).

Moiré, which is a one form of aliasing, is a problem for many studio photographers with low res backs like 20MP MFDB, so the issue is not only hypothetical.

Regarding printer resolution your 30 MP sensor would resolve around 6700 pixels horisontally. New Epsons at finest setting seem to resolve 720 pixels/inch. So if you don't print wider than 9.3" then you will never exceed the capability of a modern printer. But my standard print size is 26 inch wide.

The best MFDB is probably the Phase One IQ180 and it has 5.2 micron pixels.

Best regards
Erik

The amount of detail is not a simple function of pixel pitch. Which sensor would produce more detail, a 30MP 24x36mm sensor or a 30MP 48x72mm sensor? And when does detail become meaningless--if your viewer cannot perceive the detail or the output device unable to render the detail, what is the point of the detail?

[PierreVandevenne]

Just my 2cents. I don't think there is an "optimal" sensel size anyway. Sampling and DR are linked: it's very clear in astrophotography with very faint signals and markedly different focal lengths (real ones, not photographic equivalent). All other things being equal, there's always the unescapable trade-off between resolution and per pixel SNR (per area SNR is a different story). The range of sensel sizes we are in (say, 4-8 microns) is a good and polyvalent compromise. But yes, a camera with lots of 4 micron pixels with clean 2x2, 3x3 and 4x4 hardware binning would be really nice!

[theguywitha645d]

Hi,

Both sensors will produce the same amount of detail, but the 48x72 mm sensor will have better microcontrast. The 48x72 mm sensor will also produce a lot of fake detail unless it has a very agressive OLP filter or you stop it down to f/22-f/32. (The Leica S2 seems to cease aliasing when stopped down past f/11, your hypothetical 48x72 sensor would have almost the double pixel size so it would need stopping down two stops more).

Moiré, which is a one form of aliasing, is a problem for many studio photographers with low res backs like 20MP MFDB, so the issue is not only hypothetical.

Regarding printer resolution your 30 MP sensor would resolve around 6700 pixels horisontally. New Epsons at finest setting seem to resolve 720 pixels/inch. So if you don't print wider than 9.3" then you will never exceed the capability of a modern printer. But my standard print size is 26 inch wide.

The best MFDB is probably the Phase One IQ180 and it has 5.2 micron pixels.

Best regards
Erik

Thanks Erik, but it was more of a rhetorical statement. Also I don't think you will find an Epson will print a perfect series of line pairs at its maximum resolution and, even if it did, you would need to view it extremely closely to see it--probably with a loupe. I believe the finests setting for an Epson is 4880 dpi--720 is the lowest, but I would not confuse printer dpi with pixels.

[ErikKaffehr]

Hi,

In the latest c2ps videos Jeff Schewe says that for some pictures the 720 settings matter.

To put things in perspective I have compared 12MP APS-C with 24MP FF in A2-print, using the same setup. In one case, a picture of a mural painting, the pictures could not be told apart, but mostly the difference was subtly visible. The files from the 24 MP camera were always much better.

Some authors say that no more than 12 MP is needed, there is something to it.

On the other hand, as far as the lens outresolves the sensor you are getting more and better information with smaller pixels. If the pixels are to large fake resolution will result, which may enhance visual impression, but nevertheless be fake. With smaller pixels you can sharpen better.

This article is by a well known expert on printing: http://theonlinephotographer.typepad.com/the_online_photographer/2009/02/why-80-megapixels-just-wont-be-enough.html

Best regards
Erik

Thanks Erik, but it was more of a rhetorical statement. Also I don't think you will find an Epson will print a perfect series of line pairs at its maximum resolution and, even if it did, you would need to view it extremely closely to see it--probably with a loupe. I believe the finests setting for an Epson is 4880 dpi--720 is the lowest, but I would not confuse printer dpi with pixels.

[torger]

In the latest c2ps videos Jeff Schewe says that for some pictures the 720 settings matter.

I do my prints via a lab that has a Durst Lambda digital C-printer. I like digital C prints, unfortunately it seems like they're on their way out(?), Durst has stopped making the printers at least.

Anyway, that printer has a resolution limit of 400 ppi. It has some built-in scaling algorithm though, seems to be bicubic, which means that it smooths/smears the pixels somewhat to avoid aliasing. This smoothing is also apparent when there should be no upscaling required. This means that if I feed it exactly with 400 ppi I get a less sharp print than if I feed with with say 800 ppi. The smoothed 400 ppi is similar to effective 300 ppi. Not sure if it is possible to control the Durst to skip the smoothing at 400 ppi, I've tried to ask that to my lab but they're not really into tech nerd things so I have not got an answer (probably they don't know either and don't care to find out) :-).

[theguywitha645d]

Erik,

I agree with you.

I am not against pixel resolution. 12MP image a great and more MP can give the image more--I actually like lower rez sensors for the more 35mm filmish look or structure or rendering (it does not look like film, buy rather a limit in frequency). Our perception of images is far more complex than just reducing the problem to resolving power and pixel resolution. But likewise just going for more and more pixels has diminishing returns. When making calculations, how many significant figures do you use? Would you build a house with a ruler accurate to a micron? In images, how many pixels?

Personally, I don't know where the limit is to pixel size or to pixel resolution, at least in a simple number. I guess what I react to it a simple view the smaller the better, the more the merrier. Given a choice, I would take fewer, fatter pixels and larger sensors. But at work we do scientific imaging with 1.25MP to 4MP microscope cameras with 7-9um pixels. I have even made 24" prints from those images. So maybe I have a different perspective from most. The manufacturers will keep pushing pixels like drugs; that is what the people (and addicts) want. I am also noticing a loss of acutance in optics because they need more resolving power, so there are systemic changes that do affect the results.

My too sense.

[Sareesh Sudhakaran]

In images, how many pixels?

Personally, I don't know where the limit is to pixel size or to pixel resolution, at least in a simple number.

I did a simple amateurish calculation last week and the number I got was 5 microns - barring all else was constant and perfect. It was based on 20/8 vision (Only a tiny fraction of the population has this). Resolution of the 20/8 human eye would equate to 0.4 arc minutes at 20 feet. "Normal" vision (20/20) is 1 arc minute at 20 feet.

Since there are DSLRs out there with less than 5 microns per pixel, I suspect that it might due to resolution losses in the filtering and debayering methods adopted. Also, I would assume compression, sampling and color space all have their respective parts to play.

I can't find any data on filters on the sensors used. If I could, I might be able to get a more realistic number. But I'm just new to all this, and am still learning stuff.

[ErikKaffehr]

Hi,

I'm aware of both 0.4 arc minutes and 1 arc minute, but I don't see how you arrive at the 5 micron figure. The figure would need to include a distance and also a magnification. You never look at pixels unmagnified. For instance if you print A2 and use APS sensor image would be enlarged about 25x so a pixel of 5 micron diameter would be 125 microns.

Assuming 1000 mm viewing distance that would correspond to 0.42 minutes of arc, but you need at leas three point two resolve two. The two points and a darker point in between. That would be 1.2 arc minutes, but I'm not sure about that. Let's say 0.8 minutes of arc. Are you looking at 25 cm which is feasible it would be 3.2 minutes of arc.

But I would say it is reasonable to say that A2 prints normally are viewed at about arm lengths distance. many authors claim that 12 MP is perfectly good for A2 size prints. It also reflects my experience, although I find that in some tests I have made the difference is quite noticeable between 12 MP and 24 MP, depending on subject and conditions.

Nevertheless, there are some advantages with smaller pixels, main advantage may be that risk of aliasing is much reduced. OLP filtering can be reduced and the image will respond better to sharpening.

Best regards
Erik

I did a simple amateurish calculation last week and the number I got was 5 microns - barring all else was constant and perfect. It was based on 20/8 vision (Only a tiny fraction of the population has this). Resolution of the 20/8 human eye would equate to 0.4 arc minutes at 20 feet. "Normal" vision (20/20) is 1 arc minute at 20 feet.

Since there are DSLRs out there with less than 5 microns per pixel, I suspect that it might due to resolution losses in the filtering and debayering methods adopted. Also, I would assume compression, sampling and color space all have their respective parts to play.

I can't find any data on filters on the sensors used. If I could, I might be able to get a more realistic number. But I'm just new to all this, and am still learning stuff.