Film vs Digital - Dynamic Range

rishij · December 7, 2008

Here, let me give an example.

Say you have a 10-stop scene, and you expose perfectly to the right. Let's also say that the max voltage output of your photocell is 1024 mV. Let's break down the stops:

Stop 10: 512-1024 mV

Stop 9: 256-512 mV

Stop 8: 128-256 mV

Stop 7: 64-128 mV

Stop 6: 32-64 mV

Stop 5: 16-32 mV

Stop 4: 8-16 mV

Stop 3: 4-8 mV

Stop 2: 2-4 mV

Stop 1: 0-2 mV

If you had a 10-bit ADC, linear, then output values would correlate wonderfully to input values... that is, an input of 512 mV would give you an output of 512 (of course, in binary though), and an input of 1024 mV would give you an output of 1024.

But now let's say you have a 8-bit ADC whose voltage range is 0-1024 mV. An 8-bit ADC is capable of putting out 256 different output values. Well, let's distribute the 8 stops of the 8-bit ADC evenly across the 1024 mV range:

The following relates Voltage Input :: Bits

0-128 mV :: 1-2

128-256 mV :: 2-4

256-384 mV :: 4-8

384-512 mV :: 8-16

512-640 mV :: 16-32

640-768 mV :: 32-64

768-896 mV :: 64-128

896-1024 mV :: 128-256

A horrible distribution, I agree, but the entire 'dynamic range' of voltage is represented by this 8-bit ADC. However, it's a non-linear distribution, if I understand this correctly (if I don't, I'm sure Vijay will smack, oops I mean correct, me), because for voltages 0-128 mV, each bit increment (i.e. 1 to 2) represents 128 mV. For voltages 128-256 mV, each bit increment (i.e. between bits 2 and 4) represents 64 mV. For voltages 256-684 mV, each bit increment (i.e. between bits 4 and 8) represents 32 mV. And so on and so forth till you get to voltages 896-1024 mV, where each bit increment (between 128 and 256) in fact represents a step of 1 mV.

Now, I pose this question to any EE out there: is it possible to design such an ADC?

I don't see why not. But if not, then my entire argument here falls apart.

The other possibility is companding and then using a linear ADC, but I don't feel like explaining that right now.

-Rishi

P.S. Vijay take care of your wrist and/or get an ergonomic setup. This is not to be taken lightly. I had 1/3 of my left arm opened up to deal with nerve compression; I'd lost sensation and all muscle function of my left hand. Luckily, a year after surgery, I got most of it back.

rishij · December 7, 2008

Mauro -- any luck getting your hands on a 5D Mark II. Preliminary results show that even with the megapixel bump, noise levels are kept at or below that of the 5D. Sounds great.

But what I'm curious about is whether dynamic range is maintained and also if color reproduction is as good. The latter because they claim their color filters can each let in more of each wavelength of light, but that makes me wonder about color separation and its concomitant effect on color resolution.

Rishi

vijay_nebhrajani · December 7, 2008

Bernie - Rishi beat me to an example. A contrived one, but a valid one.

Now, I pose this question to any EE out there: is it possible to design such an ADC?

</blockquote>

Absolutely so. Thats what A-law, mu-law and other nonlinear ADCs are all about. If you can think of a "map" - a function or table that gives you a number for a voltage, you can make it. Nikon's NEFs are linear at the lower end and parabolic otherwise.

<li>When the bit depth of the ADC is greater than the DR of the recorded scene, then final DR is *not* dependent on the ADC bit depth.</li>

<li>When the bit depth of the ADC is lower than the DR of the recorded scene, the the bit depth of the ADC is DR limiting (i.e. you won't be able to capture more stops than the # of stops the ADC can represent).</li>

</ol></blockquote>

Case 2 is called clipping. We've been over that many times. Clipping is a limiting case when the signal exceeds dynamic range of the recording medium. You can have a scene DR of 0 stops - that gray wall; and overexpose (or underexpose) by 10 stops; you will clip. This has nothing to do with dynamic range of the scene even; much less the bit depth of the ADCs. You'll even clip with analog devices - such as film. You'll even clip if your bit depth was 1000 bits - because - this isn't about bit-depth. It is about the signal exceeding the DR of the recording medium. For all your talk of linearity, Bernie - have you seen the dynamic range curve from a digital camera? Here's a link: <a title="EOS 1 Ds III DR" href="http://www.dpreview.com/reviews/canoneos1dsmarkiii/page21.asp" title="EOS 1 Ds III DR">EOS 1Ds Mk III DR</a> - the curve is S-shaped. So at the ends, you don't quite get one stop change on the image file for one stop of light change. For this particular camera, the curve isn't even linear for very long - maybe the center 2-3 stops out of the ~9 stop DR. A doubling of exposure not giving you a doubling on the image file - if you know how to interpret a graph, then please do so and explain the non-linearity of the DR curves. Thanks for the warning about my wrist, Rishi. I'm going to have it looked at.

charles_wood · December 7, 2008

All in all, an entertaining thread but somewhat irrelevant to those of us that for years have been shooting color transparency film for landscape and nature images, and dealing with the limited range of transparency film stocks. Many photographers, myself included, have preferred the 'look' of Fuji Velvia vs negative film stocks, in spite of their greater dynamic range, and have learned to work around Velvia's limited dynamic range. That said, the dynamic range of most DSLRs now exceeds that of transparency film so the entire issue for my type of photography becomes moot. And carrying the scenario an additional step forward, stitching software is now so advanced that it is possible to create a large file from multiple digital frames that equal or surpass the clarity of all but the largest film formats. And HDR processing will continue to evolve and move photographic art forward. But the bottom line is that my customers don't lose sleep over how the image was captured or created. If it evokes meaning or emotion, then it has succeeded regardless of whether it arrived via film or digital, or a combination of both.

rishij · December 7, 2008

Since we're on the subject of dynamic range & the Canon 5D Mark II has been brought up, I figured I'd pose this question here... though, Mauro, if you want I could move this to a new thread.

I brought up earlier the observation of banding in extreme shadows of Canon digital images. More easily seen by lightening shadows using 'Fill Light' or whatever algorithm you choose.

Am I to understand that this banding results from variations in the amplifiers for each row when these variations exceed the signal:noise ratio (which is much more likely to occur at low signal, i.e. shadows... OR, at high ISO where the gain for each amplifier is increased)?

Here's an example of some horizontal banding (note, the image was significantly brightened in DPP):

Here's a comparison of the same area of a shot, at ISO 50 & at ISO 100, on the 5D Mark II, with Fill Light set to +100 to exaggerate the vertical banding in shadows:

Here's the link to the <a href="http://staff.washington.edu/rjsanyal/Photography/Canon5DMII_ISO50vsISO100.jpg">full-size image.</a>

And here's the banding seen in a high ISO image (ISO 25600):

One thing that confuses me though -- sometimes it seems like the direction of the banding changes from image to image, by 90 degrees (and no, not because the image was portrait vs. landscape). I'm confused...

Help? Rishi

mauro_franic · December 7, 2008

Rishi, the person who offered to bring one and participate in the test bailed out. (or returned it due to the black spots on to the right of highlights problem reported).

Looking for another person in the area.

The Sony A900 was disapointing resolving only 2700 lines per picture height. The dynamic range seems on par with most current digital though.

rishij · December 7, 2008

Am I to understand then, Vijay, that buildup of charge in the photocell is NOT linear, but with a 'toe' and 'shoulder' region? I confess, I haven't read your links yet to the photocell design, etc. Though I fully intend to!

Also, when I wrote: "2. When the bit depth of the ADC is lower than the DR of the recorded scene, the the bit depth of the ADC is DR limiting (i.e. you won't be able to capture more stops than the # of stops the ADC can represent)."

I wasn't talking about clipping. Let me be more clear: 'When the bit depth of a linear ADC is lower than the the number of stops of the recorded scene, but the max output voltage of the photocell is matched to the max input voltage of the ADC, then the bit-depth of the ADC is DR limiting UNLESS it is a non-linear (or some other design) ADC or unless you use companding (like the example of sending a microphone signal over a limited bandwidth wireless signal, etc.)

-Rishi

rishij · December 7, 2008

"due to the black spots on to the right of highlights problem reported"

</blockquote>

Sorry, don't know what you're talking about here... please elaborate? Or provide link? Thanks, Rishi

mauro_franic · December 7, 2008

Charles, "stitching software is now so advanced that it is possible to create a large file from multiple digital frames that equal or surpass the clarity of all but the largest film formats"

If you are God and can stop leaves from moving, clouds and tree leaves from changing position, control perispheral beahavior of lenses, etc, then that would be a serious option, but it's not.

"And HDR processing will continue to evolve and move photographic art forward"

Same as the above. Have you looked at a landscape shot with HDR up close? Does not work. This is only OK for web-size pictures. You may get lucky, but most likely you'd waste an outing.

"But the bottom line is that my customers don't lose sleep over how the image was captured or created."

That's right.

rishij · December 7, 2008

Charles:

Many photographers, myself included, have preferred the 'look' of Fuji Velvia vs negative film stocks, in spite of their greater dynamic range, and have learned to work around Velvia's limited dynamic range. That said, the dynamic range of most DSLRs now exceeds that of transparency film so the entire issue for my type of photography becomes moot.

</blockquote>

I couldn't have said it better. Exactly why I'm planning on ditching Velvia 50 in favor of the 5D Mark II with my holy trinity of L-series lenses.

Rishi

mauro_franic · December 7, 2008

Rishi, here this is DPReview's forum front page:

http://forums.dpreview.com/forums/forum.asp?forum=1032

Open any thread that has "black" on the title.

rishij · December 7, 2008

Interesting. The effect is explained here:

http://www.appliedcolorscience.com/black_sun.htm

I wonder if Canon normally deals with this on their sensor, or with software. If it's the latter, perhaps the RAW converter just needs to mature for the 5D Mark II.

-Rishi

mauro_franic · December 7, 2008

That explanation is just a theory (not validated yet). A fact is that many people have withdrawn their 5DII orders due to this. ... And made it harder to find a testing buddy with the 5DII.

mauro_franic · December 7, 2008

I still don't expect the 5DII to match Ektar in either Drange or resolving power. But it will certainly be a nice FF film-digital al la 2008 comparison.

helenbach · December 7, 2008

Vijay,

"For all your talk of linearity, Bernie - have you seen the dynamic range curve from a digital camera? Here's a link: EOS 1Ds Mk III DR - the curve is S-shaped. So at the ends, you don't quite get one stop change on the image file for one stop of light change. For this particular camera, the curve isn't even linear for very long - maybe the center 2-3 stops out of the ~9 stop DR. A doubling of exposure not giving you a doubling on the image file - if you know how to interpret a graph, then please do so and explain the non-linearity of the DR curves."

I hate to say this, but when I look at the graphs on that page I get the impression that I am not looking at what comes off the ADC (ie Raw) but at values after a tone curve has been applied. They don't tell you anything concrete about what is happening between the detector and the digital side of the ADC (ie analogue gain and ADC algorithm).

Best, Helen

bernardwest · December 7, 2008

Just to make sure I understand what you're saying... you claim that:

When the bit depth of the ADC is greater than the DR of the recorded scene, then final DR is *not* dependent on the ADC bit depth. When the bit depth of the ADC is lower than the DR of the recorded scene, the the bit depth of the ADC is DR limiting (i.e. you won't be able to capture more stops than the # of stops the ADC can represent).

</blockquote>

Yep, this is what I am saying. I am accepting that a non-linear ADC could change this equation, but I am claiming that the ADC's in dslrs are more-or-less linear, so these two premises will hold.

The following relates Voltage Input :: Bits

0-128 mV :: 1-2 128-256 mV :: 2-4 256-384 mV :: 4-8 384-512 mV :: 8-16 512-640 mV :: 16-32 640-768 mV :: 32-64 768-896 mV :: 64-128 896-1024 mV :: 128-256

A horrible distribution, I agree, but the entire 'dynamic range' of voltage is represented by this 8-bit ADC.

</blockquote>

Thanks for doing the example Rishi. One comment on this - Dropping one stop (i.e. 1024 -> 512) doesn't lead to a halving of histogram levels (i.e. 256-> 128; instead it leads to a 1/16th drop). As this isn't what we see happening with dslrs, we should be able to confidently say that the ADC doesn't function this way.

But this is good (Vijay too, with your link - which I will get to in a minute), we are getting to the real issue here. At least you guys have stopped saying things like 'bit depth isn't related to DR, full stop'. We see here how it can be linked, if the ADC is inputing and outputting linearly. So, as I have been trying to elucidate for a while now, it is this issue of linearity that is the CRUX of this argument. If it can be shown that the conversion is linear, then it doesn't matter what happens between scene and untouched raw data file.

So, to the question: Is it a linear conversion? In practice, when noise is thrown in the mix, it can't be 100% linear. Noise will proportionaly affect shadows more than highlights, so the 'toe' end you would expect to be non-linear. The magnitude of this non-linearity will be important. If it is only slightly non-linear, then it will behave ALMOST linearly, and the stop to bit depth ratio will remain close to 1. If it is grossly non-linear, then the stop to bit-depth ratio will vary significantly from 1. If I could be bothered, i would work it out. But the reason I won't bother is that even if it did vary greatly from a ratio of 1, the toe of the relationship is small compared to the rest of the graph.

In summary, basically the lowest signal (including noise) in the relationship will be the most non-linear. As you move up the levels (ie. increasing signal) the signal to noise ratio increases and the relationship moves closer to being perfectly linear. The real question is: "HOW close to linear is the relationship?" I say it is close enough to more-or-less directly and linearly convert stops to bits and vice versa.

Actually, I will have to think about Vijay's link a bit more, and do some more research. I'm not entirely sure what they are showing here. Without delving into this deeply at this stage, I was of the understanding that DPReview use JPG's for their DR testing. If that is the case, this graph is totally useless for this discussion. I'll look into this more.

rishij · December 7, 2008

But Bernie, how is any of this relevant if the DR of the photocell itself is, thus far in most cameras I believe, lower than the bit depth of the ADCs that are in these cameras??

EVEN IF ADCs in the cameras are linear, the fact that their bit depth is higher than the DR of the photocells (I remember reading somewhere that these photocells/amps are only capable of putting out between 2000-4000 different voltages?) means that the bit-depth of the ADCs, as long as they're as high as they are, don't affect the DR.

EXCEPT that higher bit-depth ADCs better represent shadows, and therefore most likely give you more usable DR in the shadow range.

-Rishi

bernardwest · December 7, 2008

But Bernie, how is any of this relevant if the DR of the photocell itself is, thus far in most cameras I believe, lower than the bit depth of the ADCs that are in these cameras??

</blockquote>

Rishi, you realise you have just related DR to bit-depth? I don't know if this was concious or unconcious, but you are effectively agreeing that DR and bit-depth are linked somehow (as Vijay did a thousand posts ago, and hasn't acknowledged yet). I'm not sure what your argument is here. This is how I understand it to work: photocell has a max voltage and the ADC has a max voltage. In the real world, the ADC max is greater than the photocell max (Vijay explained the engineering reasons for this somewhere above). So, the max dynamic range that the photocell/ADC combination can provide is limited by the lowest DR in the chain, which is usually the photodiode. But this doesn't affect DR that much as I explained with the 40D example. Max ADC output is 2^14=16384, but in reality it has no data in levels (approx) 13850 to 16384. But when you translate this to bits/stops, it is somewhere between 13 and 14 bits/stops. So you only lose about 0.5 of a stop in this process.

helenbach · December 7, 2008

Bernie,

"Actually, I will have to think about Vijay's link a bit more, and do some more research. I'm not entirely sure what they are showing here. Without delving into this deeply at this stage, I was of the understanding that DPReview use JPG's for their DR testing. If that is the case, this graph is totally useless for this discussion. I'll look into this more."

You must have been writing that as I made my previous post. I came to the same conclusion - those curves are not ADC output, but appear to be ADC output with a tone curve applied.

Best,

Helen

vijay_nebhrajani · December 7, 2008

Rishi, you realise you have just related DR to bit-depth?

</blockquote>

Bernie - don't you understand the English language? How many times do I have to say that the relation is incidental - i.e., based on a conscious decision by the engineers? What idiot would pair a photodiode with a huge dynamic range to an ADC that couldn't handle it? After all, even a nonlinear ADC has an upper limit to what it can handle, so if you put in a photodiode that exceeds its range, even that will clip. For the umpteenth time, I said that the relation you see is incidental , not non-existent . Read my previous posts or just search for the word "incidental" in this page.

There is no inherent relationship that bit depth has to dynamic range; the "relationship" you see is incidental: a consequence of design choices. But there is an inherent relation (i.e., a fundamental, unchanging relationship) between bit-depth and number of levels or tones or "tonal range".

bernardwest · December 7, 2008

Vijay, calm down. By continually shouting that there is no relationship between DR and Bit-depth doesn't just make it so, and it certainly won't convince me. Give me examples of why it isn't so, then you stand a chance of convincing me.

Couple of comments:

What idiot would pair a photodiode with a huge dynamic range to an ADC that couldn't handle it?

</blockquote>

You are talking about DR range of both the photodiode and the ADC. Rishi (and you earlier) specifically wrote a sentance where you equated bits and DR. Don't blame me for what you wrote. Here's what you actually wrote at 7.40pm on 5th december:

....A nonlinear ADC can encompass a larger dynamic range with fewer bits.

</blockquote>

There is no inherent relationship that bit depth has to dynamic range; the "relationship" you see is incidental:

</blockquote>

Who cares if it is incidental, accidental, on purpose, or was designed by aliens. The fact is, this IS the way the sensor electronics function, and as the relationship is more-or-less linear, we can infer all sorts of stuff from this.

imagineitkc · December 7, 2008

This helps a lot guys. I feel so much better. I trained and learned on a Cannon EOS something that was a 35mm film camera, really good at the time. I then got hooked with the medium format and fell in love. Finally, I learned how to shoot with large format. I even carted that baby around into some fields.

Anyway, I shot and developed my own film. Sometimes I sent off, but I usually printed by hand. I dropped my 35mm camera off of a pier in Boston, it worked perfectly for a year and then one day it stopped. I went digital. I had wanted to go in that direction and let’s face it, it is cheaper to not have to process all the negatives or buy all the chemicals. I wasn't in college anymore so I didn't have access to all the equipment as before and a simple apartment just doesn't have that kind of ventilation. :) But I have struggled with my camera not capturing exactly what I see. I could get it with any of my film cameras but why not digital. Was it my camera? Did I not know how to use my equipment? What is it? So, I studied and read up found out that my camera did have issues, the light sensor is a few degrees off. Which can make a difference, a huge one, but even compensating for it, it still isn't right. I’ve shot with other cameras and have taught photography, even with others, no matter what type of camera it is, the digital is different than film.

I love how easy it is to shoot, download, and order prints digitally. But I really hate that I can't capture a moment the exact way I see it. I am someday going to figure this out.

Hey, I am going to purchase a new camera. I am very, very, very much a professional photographer. Meaning this is all I do with two kids, two cats, and a husband that stays home to help my run our business. There is no income coming in other than what I photograph. That being said I need a new camera. What are your suggestions.

P.S. If I could afford a Hasselblad digital medium format I would have it. I am looking at the $3000 range or smaller, maybe a little higher. I know that some of you know a lot more about this than I do, thank you for your help.

bernardwest · December 7, 2008

You must have been writing that as I made my previous post. I came to the same conclusion - those curves are not ADC output, but appear to be ADC output with a tone curve applied.

</blockquote>

Helen, yes I wonder about this. I have heard it said numerous times on numerous forums that dpreview use jpgs for there DR tests. But I can't find anything about their methodology on their website.

helenbach · December 7, 2008

Another question - are all the ADCs used in digital cameras integer-output ADCs, or have ADCs with a floating-point output been implemented in cameras? That would delink dynamic range from bit depth even with a linear algorithm, but would still maintain the link between tonal resolution and bit depth.

Thinking about it, that would explain some of my suspicions about the behaviour of recent cameras better than a non-linear ADC.

Of course there is only a fine line between an integer logarithmic algorithm and a floating-point linear output.

Best, Helen

mauro_franic · December 7, 2008

Anyone with the 5DII in the Atlanta area yet to run a joint test?

Sign In

Film vs Digital - Dynamic Range

Recommended Posts

Link to comment

Share on other sites

Top Posters In This Topic

Popular Days

Top Posters In This Topic

Popular Days

Posted Images

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Link to comment

Share on other sites

Create an account or sign in to comment

Create an account

Sign in