Film vs Digital - Dynamic Range

[jamesnjohnson]

Adorama posted the DxOmark scores as follow:

The top 20:

 

1. Fujifilm FinePix S5 Pro (13.5)

2. Fujifilm FinePix S3 Pro (13.5)

3. Nikon D90 (12.5)

4. Sony Alpha 900 (12.3)

5. Nikon D3 (12.2)

6. Nikon D700 (12.2)

7. Nikon D300 (12)

8. Canon EOS 1DS Mark III (12)

9. Canon EOS 1D Mark III (11.7)

10. Pentax K10D (11.6)

11. Sony Alpha 350 (11.5)

12. (tied) Nikon D200 (11.5)

Nikon D40x (11.5)

13 (tied) Canon EOS 50D (11.4)

Konica Minolta Maxxum 5D (11.4)

Nikon D60 (11.4)

14 Canon EOS 1Ds Mark II (11.3)

15 Canon EOS 40D (11.3)

16 (tied) Sony Alpha 100 (11.2)

Samsung GX 20 (11.2)

 

 

So the I take it that the S5 is the best candidate for a DR comparison between the Film and Digital.

[rishij]

Robert,

 

Yeah pretty much I think.

 

Think of it this way: for any given pixel, there's a minimum and maximum voltage that the photocell can output in response to light falling on it. This determines its dynamic range.

 

For film, for any given area of film that can resolve the total, or reasonable, number of tones represented by the film (because this is what I consider the 'pixel equivalent' of film... why? I can't get into it here, but it's partially explained in that gargantuan 'is film binary?' thread), there appears to be a greater ratio between the darkest patch of silver corresponding to the brightest image detail recorded and the lightest patch of silver corresponding to the darkest image detail recorded. ADDITIONALLY, the observation/hypothesis that film response starts decaying (shows non-linearity) at high exposures (perhaps via the explanation I gave above about silver specks not wanting to grow too large b/c they start disrupting the crystal lattice of AgBr grains) results in a 'companding' phenomenon that further extends the dynamic range of film for perhaps the last 2-3 stops of information recorded.

 

Once photocells can put out a higher voltage in response to greater amounts of light than is currently implemented, or once companding can be implemented on a pixel level (photocell response decays with increasing photon exposure), the DR of digital may very well extend that of film.

 

Which is extremely exciting.

 

Rishi

[bernardwest]

<i>Vijay,

 

Why does Bernie observe that the max signal for a blown out image from a 40D is only circa 13,000? </i><p>

 

What, am I invisible? It's either because of one of two things (which have both been explained already): The photosites saturate at 13,850-odd, as that is what their full-well capacity is. This is equivalent to say 13.5 bits (settle down vijay, I know this isn't possible...). So if they put a 13 or 12 bit ADC in, then they would lose data. Obviously they need a 14 bit ADC. The other possible reason is, as I stated, digital sensors aren't truly linear in the highlights, and the photosites max out before the light response becomes too non-linear.

[bernardwest]

James... sorry you keep getting ignored. You are right, the fuji is the best DR small format dslr money can by. It has been mentioned a few times before, and right now we are in the middle of a discussion of what exaclty the DXO figure really means.

[rishij]

Oh yeah, and the DR of negative film is also extended by having layers of different sized crystals.

<p>

See here:

<br>

<img src="http://a.img-dpreview.com/news/0301/fujisuperccd/slide1.jpg">

<p>

Rishi

[rishij]

Bernie,

 

What I was saying was: if the max output voltage of the photocell is 1.5 volts, but the ADC max input is 2.0 volts which it maps to 16,384 (2^14), then why wouldn't you amplify the signal from the photocell by ~1.333 before you fed it to the ADC?

 

Vijay,

 

Yeah I also was thinking about getting the 5D Mark II. I'm curious to see what its DR will be, though I doubt it'll be more than the 1Ds Mark III, right? I still want to replace film to shoot people b/c I hate trying to shoot people on ISO 160 film that I have to expose at ISO 100 anyway. Plus I still maintain that a 12MP 5D file looks as sharp, if not sharper, than a 20MP LS-9000 scan of Portra 160NC bicubic downsized to 12MP...

 

Rishi

[vijay_nebhrajani]

I'm not trying to pick a fight with you Bernie. I know what formulae are used for calculating theoretical dynamic range or SNR of a noisy channel - this is my bread and butter stuff - these make lots of assumptions about the actual noise and generally ignore thermal dependence of noise etc.. These are only useful for making engineering decisions about design feasibility and so on. I'm merely pointing out that a theoretical value of dynamic range has absolutely no meaning for a channel. It gets drilled into communications engineers that DR and SNR must be measured to get any meaningful numbers.

 

If you didn't notice, the second chapter of Emil's paper is called "Measuring Noise".

 

Thus, for all purposes, there is only one dynamic range - the actual, measured one. This is the limit of the channel, theory be damned. The channel can't carry any more information, period.

 

Now Mauro's test actually carries out a measurement - so I can't see how his methodology is at fault here. In fact the measurement of DR is even simpler than Mauro has done. Just take a gray card, keep underexposing till you hit black and record the exposure. Then keep overexposing till you hit white and record the exposure. Divide, and voila, you have the DR.

 

Don't complicate this any further than it needs to be, please.

[jamesnjohnson]

Bernie no problem, I'm used to getting ignored (Married with children)

The above illustration is important.

This difference in sizes of of crystals that Fujifilm was going after, which is NO SURPRISE being that they make film. So it seem like a natural route for a film company to take when they make a Digital Sensor. From my understanding Fujifilm was one of the first to use a 14bit ADC.

 

IT will be a shame if fujifilm pulls 100% out of the DSLR market. Because they are on the right track.

[jamesnjohnson]

Well I feel good seeing that Rishi Sanyal used a Fujifilm illustration :-)

[bernardwest]

<i>Thus, for all purposes, there is only one dynamic range - the actual, measured one. This is the limit of the

channel, theory be damned. The channel can't carry any more information, period. < p>

 

Now Mauro's test actually carries out a measurement - so I can't see how his methodology is at fault here. In fact the

measurement of DR is even simpler than Mauro has done. Just take a gray card, keep underexposing till you hit

black and record the exposure. Then keep overexposing till you hit white and record the exposure. Divide, and voila,

you have the DR. <p>

 

Don't complicate this any further than it needs to be, please.</i><p>

 

Arrrggghh... I am having flash backs!.... You really aren't following what I am saying. Your stuck in your engineering

frame of reference and need to widen your scope. I can accept that theoretical DR can be different from actual DR,

but I wasn't talking about theoretical. Like you, I am talking about measured DR. See, we aren't in as much

disagreement as you seem to think. The part that we are disagreeing on (and that's only because you aren't thinking

clearly enough) is that USABLE DR is what is important for photographic purposes. USABLE DR is different (and

less) than your measured strict engineering DR. The reason for this is that at and just above the minimum signal

(the noise floor), a viewer can't distinguish signal from noise. I can't make it any more clearer than this. This is a

highly accepted view of dynamic range in photographic circles. Yes it differs from the strict engineering view, but if

we go by the strict engineering view, then we get nonsensical results in regards to the final image.

[vijay_nebhrajani]

<i>Rishi: What I was saying was: if the max output voltage of the photocell is 1.5 volts, but the ADC max input is 2.0 volts which it maps to 16,384 (2^14), then why wouldn't you amplify the signal from the photocell by ~1.333 before you fed it to the ADC? </i>

<p>

Plenty of reasons.

<p>

Based on the design you could need as many amplifiers as columns. Like 3000-5000 of them.

<p>

These amps would amplify the signal, but would also amplify the noise. Amplifiers can't increase dynamic range. They can only worsen it by adding their own noise.

<p>

Simple transistor amplifiers are not very linear, so to maintain linearity you'd need a good Op-Amp (operational amplifier) which is a huge circuit that consumes lots of die area and lots of power.

<p>

So it is far better to let the output voltage range of a photodetector fall where it may on the input voltage range of an ADC without worrying about whether you can hit the ADCs maximum or not. As long as the ADC has enough precision for some pre-specified design need, thats it.

[jamesnjohnson]

Link to the Adorama results for DR from Dx0mark

http://www.adorama.com/catalog.tpl?op=NewsDesk_Internal&article_num=111908-2

 

I to say that there is point when the resulting SNR numbers from a scientific/engineering test won't matter to the viewer. Need to do a what is viewable SNR test. Noise will be there but what is the threshold that need to be reach to make it an issue with the final images, pure mathematical calculations won't get you there actual imagine making will. OR better yet a combination of the two to understand what the threshold is.

 

From what I gather with this thread is that DSLR are not good when compared to film, which is far from the truth.

[vijay_nebhrajani]

<i>Bernie: The part that we are disagreeing on (and that's only because you aren't thinking clearly enough) is

that USABLE DR is what is important for photographic purposes. USABLE DR is different (and less) than your

measured strict engineering DR. The reason for this is that at and just above the minimum signal (the noise

floor), a viewer can't distinguish signal from noise.</i>

<p>

Man, Bernie - if you can measure it you can use it. You can't measure a signal that's lost in the noise floor; so

it is neither measurable nor usable. You can't measure something after it has saturated to a peak value, so that

is also neither measurable nor usable. Anything in between the noise floor and the peak saturation value is both

measurable and usable. "Measurable" doesn't exclude the noise floor, but the "theoretical" DR does. That's why

the "theoretical dynamic range" is a largely useless statistic.

<p>

So what is the difference between "measurable" and "usable" again? And why are they different?

[bernardwest]

<i>Based on the design you could need as many amplifiers as columns. Like 3000-5000 of them. <p>

 

These amps would amplify the signal, but would also amplify the noise.</i><p>

 

But what about the amp for ISO? Why can't this one be used to do it?

[jamesnjohnson]

Fix will be multiple sized photodiodes within the photosites to capture different tonal ranges without having to use amplification to boost the sensitivity. Being that different size photodiodes sensitivity changes with it size.

 

With a SuperCCD FX sensor room for more than 2 size photodiodes could be possible expanding the Dynamic Range and a possible lower SNR due to not using amplification on an extreme level.

 

Well this is a non-engineering scientific guessamation.

[bernardwest]

<i>So what is the difference between "measurable" and "usable" again? And why are they different?</i><p>

 

OK, 'usable' might not be the best word to a literal-robot such as yourself. Think of it as 'useful' perhaps, or

maybe 'distinguishable'. An engineering measure of DR is max signal / min signal (noise floor) right? So in a linear

raw file, the lowest bit contains 2 levels. The next lowest bit contains 4 levels, and so on. Now, the read noise (and

other noises as well - nb. not my forte, so go easy) swamps the lower few bits of information right? I'm not sure how

many it usually is, but it certainly swamps some of them. Now say the noise floor is set at 3 bits (8 levels). So in

the next bit you have 16 levels of information. What i am trying to say is that a viewer of an image is unable to

discern those 16 levels in any meaningful way spread in amongst 8 levels of noise. In a simplistic sense, the SNR is

2. This is too low for photographic purposes. When the signal gets higher, it gets to a point where the signal can

start to be discerned from the noise. This is the point at which your set you lower signal limit for the purposes of

photographic 'usable' dynamic range.

[bernardwest]

sorry, got to rush home; but my bit math is surely stuffed in the above example. Feel free to correct it, but it still doesn't change the substance of what I am saying.

[jamesnjohnson]

I think I understand what Bernie is saying.

Correct me if I am wrong but basically "Usable" = what is actually viewable, optimal quality, passable, feasible, used in the real world, and no longer have an effect on the printable image.

 

Where as measurable can go way beyond the "usable" threshold. It is no surprising that some engineering types will get caught up on the numbers and rather have something hit the roof because numbers are good, but all along just reaching the second floor produced usable quality and anything above that will not be seen and only can be measured.

[rishij]

James,

 

Why are you happy to see me use a FujiFilm image? Am I missing something here? Hope I don't need to post a credit... it should be obvious. Other than that, I shoot Fuji Velvia exclusively for landscapes. So I'm entirely partial to Fuji :)

 

Now, you say, "Fix will be multiple sized photodiodes within the photosites to capture different tonal ranges without having to use amplification to boost the sensitivity. Being that different size photodiodes sensitivity changes with it size."

 

Good idea. BUT, how many different sized photodiodes will you place within a given region? If you place too many, the effective pixel size becomes huge... which is no good. The advantage that film has in this regard is that, sure within on layer you can have many different sized crystals (indeed if they were all the same size, with exactly the same sensitivity, the tonal range would be quite limited), but you can have different sized crystals in different layers. Why is this important? Because you can stack a layer of less sensitive crystals directly below (or above, though that's counterintuitive and dumb) a layer of higher sensitive crystals, and some portion of light WILL pass thru. Therefore, within one 'cone of light' of a certain diameter penetrating the depth axis of film, you'll have multiple sized crystals of different sensitivities sampling this 'cone of light'... not possible with a digital sensor.

 

Unless it is, in which case I'm sure Vijay will correct me :)

 

Rishi

 

P.S. Vijay & Bernie, you guys are arguing the same thing at this point about 'measurable' vs 'usable' or whatnot. Boring. Let's get back to the topic at hand. That is, have we sufficiently explained the extended dynamic range of film? I think we have actually. This has been a great & incredibly informative discussion.

[hclim]

"By the way, Vijay, you seem to have become really calm & patient and have stopped any & all derogatory remarks/attacks... and I've been largely absent as of late -- hope the two aren't related :P"

 

This has been quite entertaining. Please carry on guys.

 

Vijay: "There are not two dynamic ranges - like a garden variety one and a usable one. There is only one - this is the actual, measured dynamic range, what you call the "usable" dynamic range. This is the measured maximum divided by the measured minimum (the min is equal to the noise floor)."

 

I understand in astronomy, CCDs are supercooled for higher sensitivities because noise will be less at lower temperatures. Does this mean if we cool a DSLR, we ought to get higher DR?

[vijay_nebhrajani]

<i>Bernie: An engineering measure of DR is max signal / min signal (noise floor) right?</i>

<p>

Yes, it is the smallest signal value that can be distinguished with certainty from the noise. We say this is the

noise floor for brevity, but in reality it is just a tiny amount more so that you can distinguish it from the noise.

<p>

The point you are making is that you want a signal that is even greater than this so that it can be distinguished

"even more certainly" from the noise for passable quality - say two or four times the "engineering minimum".

Well, that is just a subjective opinion. Your definition of "passable quality" is not my definition of "passable

quality". All we "engineering types" do is define the maximum "passable quality" achievable. This is an objective

measure so we can use it to compare <i>objectively</i> against other systems, unlike all those subjective

comparisons that essentially have little real meaning. Remember that SQF (Subjective Quality Factor) scam?

<p>

Thanks James Johnson for making "engineering types" a bad word. You just forgot that the entire world around you

was built by us engineers.

[vijay_nebhrajani]

<i>Rishi: Therefore, within one 'cone of light' of a certain diameter penetrating the depth axis of film, you'll have multiple sized crystals of different sensitivities sampling this 'cone of light'... not possible with a digital sensor.

<p>

Unless it is, in which case I'm sure Vijay will correct me :) </i>

<p>

I will indeed correct you. One word. Foveon.

[jamesnjohnson]

Rishi then another option will be an mix of the Foveon and SuperCCD sensor (Even though the Foveon is a CMOS) Foveon uses a multilayer design. So a multi-sized photodiodes on a multilayer sensor could be possible.

 

So there is a possibility to have a multilayer digital sensor.

 

To go back to the SuperCCD, there will no need to have too many photodiodes, the current SuperCCD has only two sizes and so far is unmatched in DR when it comes to DSLR. I guess Fujifilm hit the limit on the DX format, but on a FX format there could be room for 3 photodiodes for a giving Photosite. The MP will be low compared to the competition but the MP debate is an old on. DR will be unmatched.

[jamesnjohnson]

Vijay beat me, but my a minute

[vijay_nebhrajani]

<i>HC Lim: I understand in astronomy, CCDs are supercooled for higher sensitivities because noise will be less at lower temperatures. Does this mean if we cool a DSLR, we ought to get higher DR?</i>

<p>

Yes and no. Noise of the sensor will be reduced, but circuit speed will also be affected (there will be timing violations in the parlance) and eventually operation will be impossible. The battery will also have degraded performance at lower temperatures etc. There are lots of ill effects of just plonking your DSLR in the freezer so don't try it please.

<p>

Cooling just the sensor is different from cooling the entire DSLR.