Speculations on the megapixel race or "Do I need a + 32 Megapixel sensor?"

[drmikeinpdx]

"think about the terrible "pepper" grain in Canon EOS 5D compared to the smoother Canon EOS 7D."

What a coincidence!  I have a 5D and a 7D.  I think the 7D has what you might call "pepper grain" compared to the smoother tonal details of the 5D.  I hate using the 7D for portraits, but it sure is nice for action.

I am working on ways to get better images with the 7D by making changes in post-processing.  I am finding that if you shoot RAW, it makes a huge difference how you do your post processing.  If you shoot JPGs, your in-camera settings make a huge difference.

I think the lesson here is that it is very difficult to compare the image quality of two different cameras.  No matter which one you have, you will learn to process your images to get what you want.

Mike
Portland, OR

[tjshot]

TJ, great article, thank you, it was nice to get all that info clearly laid out.

In your final but on paragraph you mention the limits of the APS, but I confess that the NEX-7 has caught my eye as a replacement to yhe G12 , a camera that was a big step backwards. Reading the blogs it seems that the lenses are the let down for the NEX cameras at the moment, but maybe Zeiss or similar will react to what I suspect with be their popularity, and make some great glass.

I cancelled my order for the new Canon on reviewing the specs, because it simply doesn't provide what I need and want, a camera to get close to bettering the H2 system (with Phase P65) at 100iso. I realize on this board I'm in a huge minority who only care about peak performance at 100iso, but it is pretty much the only iso I'll shot at, and if the lights low I use a tripod.

I'm hoping the new Sony/Nikon 36mp sensor will provide what I want, fingers crossed.

For compact cameras, as for DSLRs, my choice has always been driven by max expected print size: I try to avoid interpolation as much as I can.
For me 5x7 inches print is the standard for casual shots from compact cameras, that's why I settled for a Sigma DP2x, as it sports the best sharpness and noise for that size.
NEX series cameras have great sensors and compact bodies, but they are aimed for larger prints; lenses however are pretty bulky and often compromise the portability of the system.
This is true also for other mirrorless compact systems.
What they lack is a relly good series of pancake primes in different focal lenghts.

About your specific requirements in low iso shooting, I don't think a CMOS full frame DSLR will ever get close to your actual medium format system.
The Phase One P65 is a very good digital back and a good match for quality  medium format lenses.
Quality medium format lenses generally have a peak performance at F 11; moreover DoF requirements will force you to shoot 1 - 1.5 stops down compared to 35mm, assuming equivalent focal lengths.
Following previous assumptions, top MTF 50% value for a real shooting scenario is about 55 lpm.
Top MTF values for the P65 + real quality lense would be:
Unsharpened MTF 50% 40 MTF 10% 78
Sharpened     MTF 50% 70 MTF 10% 82 with k= 0.55 (boost) and r=1 (radius)

Taking into account the size difference of diagonal for the two sensors, the ratio to 35mm full frame is 0.64, which translates into following equivalent MTF values:

Unsharpened MTF 50% 62 MTF 10% 122
Sharpened     MTF 50% 109 MTF 10% 128 with k= 0.55 (boost) and r=1 (radius)

(Warning: The above equivalence should be worth only as a rough performance comparison, as the MTF values for different sensor sizes are not directly comparable: for example when you print (or enlarge) to the same size a same image taken with two different sensors, you are really relying on different MTF zones for the same limiting lpm value in the print. It could be MTF 30% from the smaller sensor  and MTF 50% for the bigger one, which will appear cleaner (still far from the visible noise level) and  sharper (more contrast for the same detail).
It's different from comparing values for the same MTF 50% or MTF 10% levels.)

Even a 50 Mpxls  full frame sensor 35mm with the best real lense would not deliver similar values, falling short of 20 lpm in MTF 50% and 10-15 lpm in MTF 10%.
Resolution would be close, but difference in sharpness would be quite relevant.
However, SNR ratio, Dynamic range and tonal separation would still be much better with the P65 back.

So, given your typical shooting scenario, your actual digital back is your best option.

[PeterJ]

Many additional elements factor in, resulting in final image quality: A/D converter performance (12bit to 14bit seems to produce a very high boost in dynamic range, noise performance etc.), image processor, firmware etc.

Excellent summary, on the above point for some reason I got smited last time I mentioned it up but a lot of people don't realise a high-speed ADC at 14 bit resolution is pretty cutting-edge when they want the extra DR we all desire. I just checked Analog Devices and their fastest 16-bit converter costs over $US100 for 250MSPS which you'd need for 32MP x 8fps but according to their datasheet only has a little over 12 bits of effective resolution.

What you can do is us more accurate but slower ADCs in parallel which is what Canon  have done with the new X model by using 16 in parallel I believe to get a fast reading time. For comparison I'm doing a job at the moment with a delta-sigma ADC that has 22 bits effective resolution, it costs < $5 but runs at around 10 samples per second, so capturing 32MP at that rate would take over a month.

[tjshot]

Many additional elements factor in, resulting in final image quality: A/D converter performance (12bit to 14bit seems to produce a very high boost in dynamic range, noise performance etc.), image processor, firmware etc.

Excellent summary, on the above point for some reason I got smited last time I mentioned it up but a lot of people don't realise a high-speed ADC at 14 bit resolution is pretty cutting-edge when they want the extra DR we all desire. I just checked Analog Devices and their fastest 16-bit converter costs over $US100 for 250MSPS which you'd need for 32MP x 8fps but according to their datasheet only has a little over 12 bits of effective resolution.

What you can do is us more accurate but slower ADCs in parallel which is what Canon  have done with the new X model by using 16 in parallel I believe to get a fast reading time. For comparison I'm doing a job at the moment with a delta-sigma ADC that has 22 bits effective resolution, it costs < $5 but runs at around 10 samples per second, so capturing 32MP at that rate would take over a month.

I absolutely agree.
Hope Canon has followed the same "parallel processing" route shown with EOS 1D-X for the next high res body.
Frame rate would not be crucial for intended use, so maybe reducing parallel A/D units could help cutting down the costs.
I believe A/D converter performance to be crucial for coping with Dynamic Range and Noise from smaller pitch sensors; in fact it seems to be the main component involved in squeezing a clean image and better performance out of   the same sensor area.
The new Phase One IQ 180 digital back uses 16bit A/D units and it seems to pay off, at least for low ISO; I wander if the amount of light reaching a "sub 5 micron" photosite is really enough to benefit from the extended bit depth. It was for 14bit.
Probably frame rate and processing speed constraints have driven the choice for EOS 1D-X but I believe the bit depth of encoders employed in the next high res bodies, be it Canon , Nikon or Sony,  will tell us if the upgrade is profitable/effective.

[motorhead]

To try and stop progress is a waste of time and to me the ever increasing mp count is always going to be heading upwards. I am happy that this is the case, I believe the benefits of 200 to 1000mp files when we reach those heady numbers will be far smoother colour graduation and tonal qualities, not just the normal issues that others constantly mention.

I was very depressed when the 1Dx was first mentioned here, but thinking more about what Canon have said in the past about the 1Ds being radically changed gives me hope.  Just because the current 1Ds mk3 has been discontinued does not mean that Canon will ignore those professionals who need the high mp counts. I believe its just that the rapidly approaching 2012 Olympics mean that the 1D update was more urgent.

[Chewy734]

tjshot, thanks for the writeup...  I appreciate the time you took to write that article.

[tjshot]

tjshot, thanks for the writeup...  I appreciate the time you took to write that article.

Not at all!
It really was written out of personal curiosity.

[GeorgeMaciver]

Thanks for that article by Ken Rockwell, Candyman, I'm planning lowering the resolution of my RAWs next time I'm out, to test out what he says.

 

[torger]

On the print side, doesn't anyone do c-prints any longer? I find it quite nice to order large prints from the lab. I usually do a couple of smaller test prints to verify color, brightness etc and then order the final size for fine art framing. Unlike ordering inkjet prints, c-print results are very predictable, those machines are kind of fool-proof. The main drawback is that you really only can get a single glossy photo look, there is not the wide choice of papers as for inkjets. But if you like glossy c-print is one of the best.

What I'm saying is that you don't need to be a pro to do prints larger than 13x19. To me it is more common that I do prints which is larger than this format than the other way around.

Another thing is that in 30 years I may want to make prints of pictures I made this year. I don't know what print technology will exist then. So I don't really like to limit the camera sensor to commonly available print technology if you don't need to.

[tjshot]

What I'm saying is that you don't need to be a pro to do prints larger than 13x19. To me it is more common that I do prints which is larger than this format than the other way around.

Another thing is that in 30 years I may want to make prints of pictures I made this year. I don't know what print technology will exist then. So I don't really like to limit the camera sensor to commonly available print technology if you don't need to.

For me it's more a matter of optimizing investment to actual requirements: I'd rather put my money on an pro lens that will stay with me for years, than on a (relatively) short-lived camera body.
So I just choose the camera accordingly.
Moreover I see 13x19 inches from a single shot as the perfect top size for quality personal prints: archive size is still manageable and they are big enough to appreciate fine details.
If I'll ever need to print the same image larger I could still interpolate with decent results or do digital mosaics; with a bit of knowledge even a cheap pano setup can give excellent results.
Thats why I look forward to a 32 Mpxls body, as it perfectly fits my needs.
Every further increase in res. should be evaluated against noise and dynamic range performance.

[torger]

Thats why I look forward to a 32 Mpxls body, as it perfectly fits my needs.
Every further increase in res. should be evaluated against noise and dynamic range performance.

I shall say that also agree that 32 megapixels would be an excellent resolution for 5Dmk3, especially without AA filter. Above 30 it will be much about lens resolving power anyway. High contrast pixels at 32 megapixels (thanks to a sharp lens and no AA filter) could give a sharper image in print than low contrast at 45...

Already today the APS-C cameras 18 megapixels is kind of out-resolving lenses, the individual pixels are not particularly high contrast. You may be getting almost the same resolution from a 12 megapixel fullframe camera as from an 18 megapixel APS-C, depending on lens.

I don't see it as a big disadvantage to out-resolve lenses though, with bayer array it can be nice to out-resolve a little, and if you need to rotate or perspective adjust the image it can be nice to have "soft" pixels since you will not reduce image quality as much as if the pixels are supersharp. Deconvolution can also restore a bit of the resolution. However, if there is DR and noise gains to have larger pixels that will be more valuable before starting to outresolve lenses.

32 seems like a good number today, because then you both increase the resolution significantly but still have larger pixels than on the modern APS-C sensors.

[tjshot]

Thats why I look forward to a 32 Mpxls body, as it perfectly fits my needs.
Every further increase in res. should be evaluated against noise and dynamic range performance.

I shall say that also agree that 32 megapixels would be an excellent resolution for 5Dmk3, especially without AA filter. Above 30 it will be much about lens resolving power anyway. High contrast pixels at 32 megapixels (thanks to a sharp lens and no AA filter) could give a sharper image in print than low contrast at 45...

Already today the APS-C cameras 18 megapixels is kind of out-resolving lenses, the individual pixels are not particularly high contrast. You may be getting almost the same resolution from a 12 megapixel fullframe camera as from an 18 megapixel APS-C, depending on lens.

I don't see it as a big disadvantage to out-resolve lenses though, with bayer array it can be nice to out-resolve a little, and if you need to rotate or perspective adjust the image it can be nice to have "soft" pixels since you will not reduce image quality as much as if the pixels are supersharp. Deconvolution can also restore a bit of the resolution. However, if there is DR and noise gains to have larger pixels that will be more valuable before starting to outresolve lenses.

32 seems like a good number today, because then you both increase the resolution significantly but still have larger pixels than on the modern APS-C sensors.

The concept of a sensor "outresolving" a lense is prone to many caveats.
Borrowing some infos from the document attached to original post, one could assume that to capture information from a real, excellent lense up to the MTF 10% (relevant resolution) a sensor would need to resolve:

about  90 lpm at F 16
about 135 lpm at F 11
about 159 lpm at F  8
about 195 lpm at F  5.6

Even a 50 Mpxls full frame sensor, or an equivalent-pitch 18 Mpxls APS-C one (EOS 7D), with proper sharpening applied, would be limited to about 115 lpm at MTF 10%: it would fall short of 20 lpm to reach the 135 lpm provided by the lense at F11.
For F 8 and F 5.6 the limit is even higher and would require an extremely dense sensor pitch, unpractical for actual CMOS technology due to poor noise and dynamic range performance.
Even a lense of average quality, in the  F 4 to F 11 range, would probably be delivering "more information" than a 50 Mpxls sensor can actually resolve.

But the main thing to keep in mind is that limiting resolution, or Nyquist limit for a digital sensor, is not the only element to consider when evaluating performance.
A denser pixel pitch, after proper sharpening, can squeeze more information out of the same lense than a larger pitch one, even when the lense is performing below the Nyquist limit of both sensors (for example stopped down to F16 or smaller).
In other words the same lense stopped down to F16 would deliver better sharpness and resolution on a 50 Mpxls sensor than on a 36 or 21 Mpxls one, after proper sharpening is applied.
This is mainly due to the fact that a denser pitch generally allows to push sharpening further, for a similar SNR (Signal to Noise) ratio; moreover the higher Nyquist limit will allow to recover and sharpen even the high frequency information that a lower Nyquist limit would cut off. 
Thus on a 50 Mpxls sensor you can squeeze 115 lpm @ MTF 10% out of an excellent lense stopped down to F16.
The main constraint is keeping up the SNR for smaller photosites so that they can produce a clean image and consequent better sharpening; up to this day it's been possible, mainly  upgrading A/D converters to higher bit depth units, using better microlenses and optimizing the electronics.

So the whole imaging chain, i.e. lense + sensor, should be considered when evaluating performance: considering the lense only, then assuming a sensor would be "outresolving" it due to its higher Nyquist limit, is simply uncorrect.

For actual CMOS technology I believe a 32 Mpxls full frame sensor to be the optimal compromise and 50 Mpxls full frame sensor to be the practical limit ( if ever fast and effective 16 bit A/D converters can be introduced).

[torger]

The concept of a sensor "outresolving" a lense is prone to many caveats.
Borrowing some infos from the document attached to original post, one could assume that to capture information from a real, excellent lense up to the MTF 10% (relevant resolution) a sensor would need to resolve:

Yes "outresolving" is a simplified way to see it. Lens resolution is not a gating limit, even if you improve a factor in the chain that is not the largest "blur factor" the resolution will improve. However, you start to see diminishing returns when increasing sensor resolution over some limit, and how far you think it is meaningful to go is a matter of taste. With 18 megapixels on APS-C I think the point of diminishing returns has been passed, so I don't see it meaningful to increase it further.

My guess is that it is not meaningful to have such high resolution on the sensor that you can resolve down to MTF10% of the best part (center) of the best lenses. If combined with some clever deconvolution sharpening it perhaps can be, but unprocessed no.

It would be interesting to make a practical test with state of the art sharpening techniques and prints etc to find out how high resolution you need from a sensor relative your lens. I think the result would be that you need considerably less sensor resolution than the theoretical max.

[Edwin Herdman]

I'm trying to work through the attached PDF article, but the sans serif Courier New font makes it unreadable.  Next time please use a serif font, for the love of my eyes.

[tjshot]

I'm trying to work through the attached PDF article, but the sans serif Courier New font makes it unreadable.  Next time please use a serif font, for the love of my eyes.

Sorry!
I'll keep it in mind for the follow up I'll be posting in a few days, about sensor performance for a given print dpi and size.