Home · Register · Join Upload & Sell

Moderated by: Fred Miranda
Username  

  New fredmiranda.com Mobile Site
  New Feature: SMS Notification alert
  New Feature: Buy & Sell Watchlist
  

FM Forums | Leica & Alternative Gear | Join Upload & Sell

1       2       3              6      
7
       8       end
  

Archive 2013 · More Mpix or perfect per-pixel quality?

  
 
kwalsh
Offline
• • • •
Upload & Sell: Off
p.7 #1 · p.7 #1 · More Mpix or perfect per-pixel quality?


Paul Gardner wrote:
Without going into the math, 14 bits does not have enough bits left on the LSB to handle the dark end of the well depth.


You should do the math, your assertion is completely wrong. As my high school math teacher was fond of saying "intuition sucks" and you appear to be using intuition instead of actual calculation.


Why do you think that MF is 16 bit?


Marketing for one. Second their read out is really slow so there isn't any particular disadvantage in throwing a 16-bit ADC on there. But it doesn't do anything to improve the signal. And you can demonstrate this yourself by rounding any MF RAW file to 14-bit or even 12-bit and discover there is no difference in the result.

Anyway, I suggest you do your homework on both quantization noise and Poisson statistics as well as take sometime to actually understand what the performance characteristics are for MF sensors.

Ken



Mar 30, 2013 at 01:45 PM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.7 #2 · p.7 #2 · More Mpix or perfect per-pixel quality?


RustyBug wrote:


Are the multiple (connotative/denotative) definitions for DR?

When I think of DR, I think about range of recordable information @ dark - light and the number of stops that range between these two extremes. Yet, much of the discussion is concerning noise and degrees of refinement/distinction/tonal separation within the range of that captured.


The two issues are closely tied together: sensitivity to tiny amounts of light compared to to the maximum saturation point (well depth) of the sensor both determines the dynamic range (range between the largest signal before clipping and smallest above noise), and also the finest distinction you can make between nearby brighter values. Dividing the scale between the darkest and brightest measurable point into finer intervals than the noise level in the darkest regions doesn't help tonal separation, because the same noise that limits DR into the shadows is present at the same or larger magnitude in mid-tones. A system with 12 stops of dynamic range will not benefit from dividing that range into finer intervals described by a 14-bit readout --- at the very most, it will need 12 bits to adequately describe meaningfully measurable tonal separations; in most cases, it could use even less bits representing non-linearly-spaced values (though 12-bit linear is probably the "easiest" solution for RAW storage).

Note, however, a system with good tonal separation at mid and bright values doesn't necessarily have better dynamic range than one with worse tonal separation.

Numerical examples:
a) a sensor with a "perfect" readout (no noise besides statistical shot noise) and a well depth of 2^12 = 4096 electrons has a dynamic range of 12 stops (from 1 to 4096 electrons). At "mid grey" 25% of the max signal, the noise as a fraction of the signal is sqrt(0.25*2^12)/(0.25*2^12) = 2^-5.

b) a sensor with 2^4 = 16 electrons readout noise and a well depth of 2^16: this sensor has the same 12 bit dynamic range (2^16/2^4 = 2^12). However, the mid tonal separation is better: the fractional noise at 25% max is sqrt(0.25*2^16 + 2^4)/(0.25*2^16) ~ 2^-6 (twice as good as above).

So, deep well depth (a necessary factor for dynamic range, combined with low readout noise) also improves tonal separation.

Edited on Mar 30, 2013 at 05:25 PM · View previous versions



Mar 30, 2013 at 05:06 PM
zhangyue
Online
• • • • •
Upload & Sell: On
p.7 #3 · p.7 #3 · More Mpix or perfect per-pixel quality?


One thing, I want to add for 14 bit DR sensor, there might be benefit to process with higher bit DAC to achieve truly 14 bit performance. People do this all the time with CD player which is red book 16 bit but process with 20bit and even 24 bit to implement digital cut off filter easier. Admittly, It is not purely for performance, but just industry standard, and consumer do chasing number/spec thinking 20/24 bit must be better than 16 bit.

I can see the reason for 18bit to achieve real 16 bit performance but not 20/24 bit.

For example, a state of art 24 bit dac can achieve close to 130dB DR a little short of 22 bit performance.

From cell phone, sorry for typo.


Edited on Mar 30, 2013 at 06:51 PM · View previous versions



Mar 30, 2013 at 05:24 PM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.7 #4 · p.7 #4 · More Mpix or perfect per-pixel quality?


zhangyue wrote:
One thing, I want to add for 14 bit DR sensor, there might be benefit to process with higher bit DAC to achieve truly 14 bit performance. People do this all the time with CD player which is red book 16 bit but process with 20bit and even 24 bit to implement digital cut off filter easier.


Audio recording applications can go "overkill" on sampling rate and bit depth because these days it's trivial to build electronic circuits that can measure a much wider range of frequencies than human hearing can detect, and a much wider dynamic range of sound than will ever occur in "real" recording situations.

With cameras, however, the bit depth is limited by the hardware capabilities: we don't have sensors that can "really" capture 18 stops of dynamic range, and just limit them down to 12 to save space. And, especially since people can pixel-peep images, there's always an incentive to push *way beyond* what the naked eye can take in all at once (while audio recordings only need to be good enough to sound realistic to a human listening with their human ears, not the super-sensitive-bat-hearing equivalent of zooming in on a tiny portion of an image frame).

As for processing images at higher bit depth after storage to avoid round-off errors, you're probably already doing that. For example, a lot of the image processing on Apple platforms uses Apple's "Core Image" routines, which internally work with massively-overkill 3x32-bit floating point data representations, before rendering back to a lower bit depth final product.



Mar 30, 2013 at 05:35 PM
philip_pj
Offline
• • • • •
Upload & Sell: On
p.7 #5 · p.7 #5 · More Mpix or perfect per-pixel quality?


14 bit...the Sony folks used to believe that 12 bit worked just as well with the a900, had all the papers to 'prove' it...but the new 24Mp sensor made believers of many. I am very much a lay person in this however, just pointing out they are heading in that direction, and people always want more: 12>14>16.


Mar 30, 2013 at 06:09 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.7 #6 · p.7 #6 · More Mpix or perfect per-pixel quality?


Well, lets compare a relatively new "16 bit" medium format camera, the Hasselblad H4D-40 used with the HC150mm/3.2 lens with a 14-bit D800 + Sigma 105mm.

Images at base ISO in ProPhoto studio flash lights. ISO100 for both cameras. Nikon shot at F8.0. Hassy shot at F11 with twice the strobe output, to equalize DoF. Full scene looks like this.


To see that neither camera got an unfair advantage, let's look at the raw file histograms. Here you can also see evidence of the file bit depth, Nikon goes 0-16383 (14-bit) and the Hassy goes to 65535 (16-bit). You can also see that we didn't in any way underexpose the images, this is a reasonably well exposed scene with just enough headroom to cover all specular highlights.


And then compare some really tortured crops (files developed with some shadow lift, and then made as equal as possible by manual curves in PS)


...and guess which side represents the Hassy 16 bit file...? :/

The story is the same all over the image field. The 14-bit Nikon file has better tonal resolution AND less noise no matter where you look, and no matter how you look at it.
In fact the D800 would walk all over the Hassy even if I set it on saving 12-bit raws. It would still have a bigger DR, and better tone resolution.

(The small bottle isn't representative of sharpness in the Hasselblad, I suspect that the 150mm lens is slightly decentered. Left side is sharper than the right side in all the images. Look at tonal resolution and noise, not sharpness!)



Mar 30, 2013 at 07:30 PM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.7 #7 · p.7 #7 · More Mpix or perfect per-pixel quality?


theSuede wrote:
The story is the same all over the image field. The 14-bit Nikon file has better tonal resolution AND less noise no matter where you look, and no matter how you look at it.


(assuming the Nikon is on the left in the crops, correct?)
Thanks for the comparison --- I don't get the chance to pixel-peep MFDB results very often. Do you think the Hassy sensor shows some redeeming qualities in other aspects not visible here (aside from lens selection differences), or is it generally this badly outclassed by the D800?



Mar 30, 2013 at 07:38 PM
carstenw
Offline
• • • • • •
Upload & Sell: Off
p.7 #8 · p.7 #8 · More Mpix or perfect per-pixel quality?


I suspect it has more accurate colours, especially skin tones.


Mar 30, 2013 at 09:00 PM
Mescalamba
Offline
• • • •
Upload & Sell: Off
p.7 #9 · p.7 #9 · More Mpix or perfect per-pixel quality?


I was always under impression that main point for 16/14bits are colors, not DR.

If Im correct, sensor will show same amount of noise no matter if it has 12 or 16 bit ADC.

Otherwise more noise on Hasselblad side might be result of CCD vs CMOS. CCD doesnt take any kind of exposure manipulation nicely.. Owners of M8 or M9 probably know that better than others.



Mar 30, 2013 at 10:07 PM
Paul Gardner
Offline
• • •
Upload & Sell: Off
p.7 #10 · p.7 #10 · More Mpix or perfect per-pixel quality?


Back in the 90s I worked for a company that was doing signal extraction 6db below the noise floor, so even if you don't see signal, that does not mean there is not usable signal existing. BUT first you have to capture it. I.E. 18-20 bit ADC. Then use DSP technics to recover. If it doesn't matter to you, just use jpg "8 bits is enough"


Mar 30, 2013 at 11:18 PM
sebboh
Offline
• • • • • •
Upload & Sell: Off
p.7 #11 · p.7 #11 · More Mpix or perfect per-pixel quality?


Paul Gardner wrote:
Back in the 90s I worked for a company that was doing signal extraction 6db below the noise floor, so even if you don't see signal, that does not mean there is not usable signal existing. BUT first you have to capture it. I.E. 18-20 bit ADC. Then use DSP technics to recover. If it doesn't matter to you, just use jpg "8 bits is enough"


um, what was their definition/measurement of the noise floor and is it at all analogous to this?



Mar 30, 2013 at 11:34 PM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.7 #12 · p.7 #12 · More Mpix or perfect per-pixel quality?


Paul Gardner wrote:
Back in the 90s I worked for a company that was doing signal extraction 6db below the noise floor, so even if you don't see signal, that does not mean there is not usable signal existing. BUT first you have to capture it. I.E. 18-20 bit ADC.


You can "recover" a constant signal that's below the noise floor in one sample by, e.g., averaging over many samples. However, an important feature is that you don't have to measure each sample with a super-fine-resolution ADC that could "see" the tiny signal in a single sample. When you combine the samples together for averaging, you're effectively increasing the bit depth of the measurement: combine two n-bit pixels together, and the two-pixel sum has (n+1)-bit bit depth (and cancels uncorrelated noise components by a factor of ~sqrt(2)). There's nothing to be gained by digitizing way finer than the noise level on each pixel --- you can just as easily "recover" below-noise-floor signals by multi-pixel averaging without an 18-bit ADC for each pixel.



Mar 31, 2013 at 12:01 AM
zhangyue
Online
• • • • •
Upload & Sell: On
p.7 #13 · p.7 #13 · More Mpix or perfect per-pixel quality?


mpmendenhall wrote:
(assuming the Nikon is on the left in the crops, correct?)
Thanks for the comparison --- I don't get the chance to pixel-peep MFDB results very often. Do you think the Hassy sensor shows some redeeming qualities in other aspects not visible here (aside from lens selection differences), or is it generally this badly outclassed by the D800?

The comparison really show that bottle neck is sensor itself but not how many bit of DAC. Looks like 16 bit is really overkill consider if it is true that 12 bit raw D800 can still outperform Hassy.

I wonder if Sony sensor has really DR of 14 bits, having 16 bit DAC should make more sense, but certainly not the case with Hassy here.

As for recover signal below noise, delta sigma ADC can do that as well, to re shaping noise out of interesting band to reduce in band noise, but I am not really seeing it has real advantage here than market hype.




Mar 31, 2013 at 12:28 AM
zhangyue
Online
• • • • •
Upload & Sell: On
p.7 #14 · p.7 #14 · More Mpix or perfect per-pixel quality?


mpmendenhall wrote:
Audio recording applications can go "overkill" on sampling rate and bit depth because these days it's trivial to build electronic circuits that can measure a much wider range of frequencies than human hearing can detect, and a much wider dynamic range of sound than will ever occur in "real" recording


A bit off topic. But for audio, there is theory that people can feel freq after 20k and below 20hz, but nobody can really verify that.

Since it almost became a industry standard that 24bit 192khz, everybody do that. But obviously, there is no real advantage compare say 18bit DAC can have true 16 bit red book performance. SACD and DVD audio both claim higher performance than CD, but nobody can survive a blind test. And sure they both died even earlier than old redbook format, leave with tens of 24 bits high performance DAC floating around.


Edited on Mar 31, 2013 at 08:26 AM · View previous versions



Mar 31, 2013 at 12:49 AM
kwalsh
Offline
• • • •
Upload & Sell: Off
p.7 #15 · p.7 #15 · More Mpix or perfect per-pixel quality?


Paul Gardner wrote:
Back in the 90s I worked for a company that was doing signal extraction 6db below the noise floor, so even if you don't see signal, that does not mean there is not usable signal existing. BUT first you have to capture it. I.E. 18-20 bit ADC. Then use DSP technics to recover.


And wrong again. Recovering signals below the noise floor does not require any additional bit depth of the sampling. If the quantization noise is slightly below the noise floor you are done and do not require any more bits to do recovery below the noise floor. You obviously don't really know what you are talking about here. Google "noise spectral density", "process gain" and "spread spectrum" if you want to learn what the company you were working for actually did and how it worked.



Mar 31, 2013 at 05:49 AM
theSuede
Offline
• • • •
Upload & Sell: Off
p.7 #16 · p.7 #16 · More Mpix or perfect per-pixel quality?


The quantization noise in a for the application typical 14-bit AD converter in a camera has got one (Nikon D7000, D800, Pentax K5-2), in many cases two (D7100, newer Oly and Pana constructions, the best Canon cameras) - and sometimes even three (Canon 1Dx, Canon 7D) bits more resolution than the best signal the sensor can provide.

As I tried to show, a 12-bit signal from the D800 would still better in tonal resolution and shadow noise than the 16-bit medium format top of the range models from 2012. This is a question of how well you treat the signal (how small the amounts of electronic noise power you add) BEFORE the signal goes into the AD.

And yes, the Nikon is to the left.

As long as I stay in the studio, or shoot in natural sunlight the Hasselblad color (and also earlier the P45+ we had) are easier to manage. They're not "more accurate" - in fact they're LESS accurate - but they have got a very well defined R-G difference, something most newer cameras don't (D5200 and D7100 might show a trend reversal here to my delight...).

This helps in getting more (not necessarily "better" or "more accurate") hue resolution in the green-yellow-orange-warm red color range. Unfortunately it also really screws up purples and deep blue, and it makes profiling colors for lights like fluorescents a total nightmare, physically impossible.



Mar 31, 2013 at 08:41 AM
RustyBug
Online
• • • • • • •
Upload & Sell: On
p.7 #17 · p.7 #17 · More Mpix or perfect per-pixel quality?


Thanks Michael (et al),

I think I followed most of that conceptually, but could you expand a bit on your comment @ deep well depth impacting DR vs. tonal separation. And what do you consider to be optimal/point of constraint/limitation/diminishing return regarding well depth?




Mar 31, 2013 at 11:22 AM
mpmendenhall
Offline
• • • •
Upload & Sell: Off
p.7 #18 · p.7 #18 · More Mpix or perfect per-pixel quality?


RustyBug wrote:
I think I followed most of that conceptually, but could you expand a bit on your comment @ deep well depth impacting DR vs. tonal separation. And what do you consider to be optimal/point of constraint/limitation/diminishing return regarding well depth?


I'm just working from simple first-principles without detailed real-sensor technical knowledge; someone like theSuede or kwalsh might have more actual-technology-based information to share. But the basic picture is this: sensor wells are like "buckets" that you can fill up to some maximum number of photons. The bigger the bucket, the more dynamic range you can get between the smallest signal you can distinguish (1 photon in the best case; a handful of photons with sensor readout noise), and the full-bucket count. At the same time, the more full your photon buckets, the easier it is to distinguish fine relative gradations: the 10% difference between 100 photons and 110 photons is hard to distinguish from the sqrt(100)=10 photon shot noise, but the same 10% difference between 400 and 440 photons is twice the sqrt(400)=20 shot noise. If you have gigantic buckets, then you can have plenty of photons in the mid- and shadow-tone regions (so fine gradations are better distinguished from noise), while still being able to count even larger numbers for the highlights.

IMO, if you have "unlimited" light available to work with, then "more is better" for well depth. In situations where you need to crank up the ISO to get the shot, then you're already DR/gradation limited by photon-counting physics (you don't need deeper wells if you can't even fill up a shallower one).

You should be somewhat able to judge for yourself how much even more well depth than what you have would be desirable: when you're shooting at base ISO and setting exposure to keep the brightest parts of the scene just below clipping, how much does the noise and/or lack of shadow detail bother you? With modern sensors, images usually look pretty great, so you may already be well past your own "point of diminishing returns" if you're already happy with the results you're getting. Taking advantage of deeper wells would require exposing at lower equivalent ISO settings (ISO50, 25, 10...) to fill the camera's bigger "photon buckets" with more light (and thus be sensitive to finer tonal gradations, and have more DR between the brightest and darkest signals).

You can simulate the effect of having deeper wells on your camera by overexposing an image by a stop or two, then pulling back the exposure in post. Ignore the badly blown highlights (this is what having real deeper wells would avoid), but the below-clipping mid and shadow tones will be as if shot with a deeper-welled camera (doubling the well depth for each stop overexposure). Compare to a correctly-exposed-as-shot image to see how much you care about the improvements.



Mar 31, 2013 at 02:11 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.7 #19 · p.7 #19 · More Mpix or perfect per-pixel quality?


Well depth is at the moment mainly limited by (the need for) high ISO performance.

Having a very large well improves lowest available ISO setting, but it also makes measuring out a very small charge from it - accurately - more difficult. It's like trying to accurately measure two teaspoons of salt by using a gallon container. Not as daft a comparison as you would think, at ISO6400 you're only using 1/64 of the FWC (if base ISO is 100, 6400/100 = 64).

For tonal separation, what you want is low electronic noise and PRNU, and high QE. Electronic noise and Patterned Noise (PRNU) adds in unnecessary noise that wasn't there in the image from the start. QE is conversion rate from light to electricity, and you obviously want this to be as high as possible. Light in itself contains an error margin (shot noise) that is sqrt(light energy). More light converted = error margin gets lower.

PRNU is historically where CCD sensors were better, but that balance of performance shifted some years ago. Today, only CCD's with MIL-spec after-electronics have lower PRNU than even pretty standard commercial grade CMOS.



Mar 31, 2013 at 07:49 PM
RustyBug
Online
• • • • • • •
Upload & Sell: On
p.7 #20 · p.7 #20 · More Mpix or perfect per-pixel quality?


So it kinda sounds like the math involved renders well depth attribute diametrically opposed relative to low vs. high ISO ... is that a close conceptual paraphrase?



Okay, from a practical standpoint ... lets say I want to capture a scene with 12 EV DR ... and I want to see detail throughout the spectrum, but tonal transitions are not primary consideration, total range is.

Contrast that with a scene that only contains say 4 EV DR, but I want the tonal transitions to be uber-smooth, since the total range is rather narrow.
(Does it matter if this is weighted toward shadows vs. highlights?)


What approach(es) @ well depth / pixels size / etc. sets up either scenario for optimal results?
(assuming single image capture only)



Apr 01, 2013 at 11:57 AM
1       2       3              6      
7
       8       end




FM Forums | Leica & Alternative Gear | Join Upload & Sell

1       2       3              6      
7
       8       end
    
 

You are not logged in. Login or Register

Username       Or Reset password



This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.