Home · Register · Search · View Winners · Software · Hosting · Software · Join Upload & Sell

Moderated by: Fred Miranda
Username   Password

  New fredmiranda.com Mobile Site
  

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

1       2      
3
       4       end
  

Dynamic Range vs. Tonal Range
  
 
philip_pj
Offline
• • • •
Upload & Sell: Off
p.3 #1 · Dynamic Range vs. Tonal Range


And while understanding dawns slowly, thank you both very much.


Apr 23, 2013 at 11:59 PM
alundeb
Offline
• • • •
Upload & Sell: Off
p.3 #2 · Dynamic Range vs. Tonal Range


zhangyue wrote:
but I am not quite sure that once we reach mid or high tonal range, the step still is defined by this noise level? We can amplify the signal, though, without change S to N ratio, but amplified signal can be divided into more bits/step to process in digital domain to achieve finer tonal response.



The formal definition of Tonal Range is that the noise level defines the step size. Being able to divide the signal into finer steps is not the same as more tonal range. Often when Tonal Range is mentioned it texts, it is said to be the number of gray tones that can be represented. Omitting the noise criterion is causing an initial, intuitive, false impression that Tonal Range is closer related to quantization (tonal resolution) than to SNR.



Apr 24, 2013 at 06:57 AM
zhangyue
Offline
• • • •
Upload & Sell: Off
p.3 #3 · Dynamic Range vs. Tonal Range


alundeb wrote:
The formal definition of Tonal Range is that the noise level defines the step size. Being able to divide the signal into finer steps is not the same as more tonal range. Often when Tonal Range is mentioned it texts, it is said to be the number of gray tones that can be represented.



Up to now, I never know what is the definition of Tonal range. All of above as I said is from my understanding. I was trying to learn it as planed but never be able to have time to do so. I must admit I am not knowledgeable enough about PS, print, Gamma, display, etc...

If I know the definition, there may not be any thinking process here. I am open to any correction as I said. but I would not limit myself to 'definition' itself.

Being able to divide signal into finer steps give freedom to post process digitally that we do all the time. We are trying to map 13 bit DR image to 8bit,(though I wonder if this/8bit gonna be the standard forever, 14bits ADC might be enough or having only small benefit, but is it future proof? ) doing that, remap tonal step by 'fill light', pull back 'black', increase 'exposure', 'brightness' etc.. we do treat each digital image by applying non linear gain in digital domain. Having finer step giving us freedom to do so and potentially can do better job with more advanced RAW converter or display which can handle finer tonal step than 256.

Why we have to stick with linear(the same) tonal step? can't we have finer step at higher grey zone? anybody can answer this?

Quantization noise obvious have significant impact on Low DR image than high DR image. Consider we do have option to keep finer step in ADC now, I see no reason we want limit ourselves by truncating the signal.

alundeb wrote:
Omitting the noise criterion is causing an initial, intuitive, false impression that Tonal Range is closer related to quantization (tonal resolution) than to SNR.


It is not about closer or further related to anything but do we really gain anything here by having more bit ADC in the loop. I was in the other camp before, but go through this thinking make me want to change and offer the topic to discuss.

The key to me is not do we violate the definition or not, but do we really omit things by default thinking minimal tonal step is the noise level, and discard useful information now just because our display, printer, or even raw converter can't maximize this benefit now. (though the benefit is already here as discussed.)

I want to use again the same example above, 4 bit DR with 2^8 resolution/step. I can potentially apply tone curve to those 2^8 step individually (if I have a advanced Raw converter with Super CPU) to any point I like to create tonal range curve I want. 2^4 step won't give me same performance or freedom.













Apr 24, 2013 at 08:15 AM
theSuede
Offline
• • • •
Upload & Sell: Off
p.3 #4 · Dynamic Range vs. Tonal Range


There are, as I said to Rusty two parts to all individual points in this.

The measurement of light - this is almost always in linear range, since there's a lack of exponential input type ADC's. Which is a shame, since linear is a thorough waste of both space and processing time when both light physics and human vision sensitivity are more closely related to an exponential curve.

The recording format - which in most raw's are in linear format, but some compressed raw formats use exponential range compression (Nikon, Sony). This is converted back to linear before ant raw processing is done.
...............

One has to be aware of this difference - which dataset am I referring to? The linear format input or the range of values that we convert that range into?

For linear format, we NEED 14 bits to be able to get 14Ev input range. Because that's how ADC's work. That most of the bit-resolution is totally wasted further up in brighter areas doesn't matter, we HAVE to use this ADC resolution.

In later stages after the measurement has been done, we can start to fit the input to a more compact and less wasteful format, like with an exponential curve (gamma). The exponential curve has to be a good fit for the actual input tone resolution in both the darkest parts and the brighter parts of the image, and the exponential of about "two" works very well. That's why both sRGB and Adobe RGB uses a gamma of ~2.2.



Apr 24, 2013 at 02:46 PM
RustyBug
Offline
• • • • • •
Upload & Sell: On
p.3 #5 · Dynamic Range vs. Tonal Range




With a gamma of 2.2 is being applied, does a gamma adjustment to .45 (1/2.2) approximate the original linear data?



Apr 24, 2013 at 03:01 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.3 #6 · Dynamic Range vs. Tonal Range


Yes, if the gamma is the ONLY thing that's been applied to the image... As in the case with Sony/Nikon compressed raw.

Usually you also have an S-curve contrast, some color-transformations and other stuff in between too. And since you can't know in which order those non-linear transforms have been applied, they are not reversible.

That's why accurate white-balance and other such operations don't work in jpg's, they've been through quite a lot of non-linear operations.



Apr 24, 2013 at 03:43 PM
alundeb
Offline
• • • •
Upload & Sell: Off
p.3 #7 · Dynamic Range vs. Tonal Range


zhangyue wrote:
Up to now, I never know what is the definition of Tonal range. All of above as I said is from my understanding. I was trying to learn it as planed but never be able to have time to do so. I must admit I am not knowledgeable enough about PS, print, Gamma, display, etc...

If I know the definition, there may not be any thinking process here. I am open to any correction as I said. but I would not limit myself to 'definition' itself.

Being able to divide signal into finer steps give freedom to post process digitally that
...Show more

The definition actually also looks to the A/D quantization, and it is a very precise description of what you want.
The step size at any point is defined as "The noise OR the quantization step, whichever is larger".

This is the definition used by DXOmark.com, and it says exactly the same as the words Joakim uses here.







The tonal range can never be greater than the number of quantization steps, so if you use a shallow bit depth you do get a small Tonal Range, also according to the definition.

Now I am not sure if you are talking about the very coarse quantization in a gamma corrected 8 bit image file, or in a raw file with 12-14 bits, or a theorethical case. If you are working in a very small bit space, you of course get problems with posterization, as when you try to edit 8 bit images, but nobody who is discussing Tonal Range here would even think about doing that?

It is also easy to think, since the SNR get better, that noise gets lower as you move up in luminosity. The opposite is the case. At sensor saturation, the noise is enormous compared to what it is in the deep shadows, because the photon shot noise is the square root of the number of photons. Normalized to the same scale: the more light, the more noise, and the larger the TR steps get.



Apr 24, 2013 at 04:00 PM
RustyBug
Offline
• • • • • •
Upload & Sell: On
p.3 #8 · Dynamic Range vs. Tonal Range


I was mostly good till that last paragraph ...

Gotcha @ unkowns for reverse engineering.



Apr 24, 2013 at 04:33 PM
AhamB
Offline
• • • •
Upload & Sell: On
p.3 #9 · Dynamic Range vs. Tonal Range


zhangyue wrote:
Being able to divide signal into finer steps give freedom to post process digitally that we do all the time. We are trying to map 13 bit DR image to 8bit,(though I wonder if this/8bit gonna be the standard forever, 14bits ADC might be enough or having only small benefit, but is it future proof? ) doing that, remap tonal step by 'fill light', pull back 'black', increase 'exposure', 'brightness' etc.. we do treat each digital image by applying non linear gain in digital domain. Having finer step giving us freedom to do so and potentially can do better job
...Show more

If you're going to do extensive PP it's always a good idea to export your RAW files as 16 bit PSD/TIF and only convert to 8 bit when saving it in the final format (for print/web display/etc.). I think RAW software work at 16 bits internally anyway (or even 32 bits floating point in the case of Iridient RAW Developer) so it's good to continue that in your Photoshop operations to minimize posterization.



Apr 24, 2013 at 04:41 PM
RustyBug
Offline
• • • • • •
Upload & Sell: On
p.3 #10 · Dynamic Range vs. Tonal Range


BTW ... to those contributing ... thanks for the time & effort you're sharing with us in the manner that you are. Good FM'ers are appreciated, yet too rarely expressed for being such.

Thanks again.



Apr 24, 2013 at 04:47 PM
 

Search in Used Dept. 



zhangyue
Offline
• • • •
Upload & Sell: Off
p.3 #11 · Dynamic Range vs. Tonal Range


AhamB wrote:
If you're going to do extensive PP it's always a good idea to export your RAW files as 16 bit PSD/TIF and only convert to 8 bit when saving it in the final format (for print/web display/etc.). I think RAW software work at 16 bits internally anyway (or even 32 bits floating point in the case of Iridient RAW Developer) so it's good to continue that in your Photoshop operations to minimize posterization.


But with limited bit at capture? I don't know how much benefit we get convert a 4 bits file to 16 bits and do PP there?



Apr 24, 2013 at 04:57 PM
zhangyue
Offline
• • • •
Upload & Sell: Off
p.3 #12 · Dynamic Range vs. Tonal Range


alundeb wrote:
The definition actually also looks to the A/D quantization, and it is a very precise description of what you want.
The step size at any point is defined as "The noise OR the quantization step, whichever is larger".


The exact point I was trying to make that once the signal above noise level, we can use more bits ADC to reduce quantization noise, so we have finer tonal step above deep black.

The question I have is do we really need each tonal step to be equal, we can let deep black have less step or the same small step (but no different black level)


This is the definition used by DXOmark.com, and it says exactly the same as the words Joakim uses here.

The tonal range can never be greater than the number of quantization steps, so if you use a shallow bit depth you do get a small Tonal Range, also according to the definition.


So looks like we are talking about same thing here that you agree have more bits ADC will have smaller tonal step.


Now I am not sure if you are talking about the very coarse quantization in a gamma corrected 8 bit image file, or in a raw file with 12-14 bits, or a theorethical case. If you are working in a very small bit space, you of course get problems with posterization, as when you try to edit 8 bit images, but nobody who is discussing Tonal Range here would even think about doing that?


Now, I think either I am not clear or you do not understand me correctly. Of course I am talking about raw file. That is all the talking I am try to do


It is also easy to think, since the SNR get better, that noise gets lower as you move up in luminosity. The opposite is the case. At sensor saturation, the noise is enormous compared to what it is in the deep shadows, because the photon shot noise is the square root of the number of photons. Normalized to the same scale: the more light, the more noise, and the larger the TR steps get.


Here, you lost me. Can I have some background information about noise source here: shot noise is from photons/sensor, so you mean more exposure will not only amplify signal, but also noise, so we are not change SN ratio here. Read noise is AA conversion, which is fixed, so more exposure/signal will have better SN ratio here.
Now go back to my original post that Now signal is there and signal level can be anywhere depend on DR of image. Having more bits ADC have two effect:

1. With low DR image, most bit might get wasted
2. but the low DR image also have more step which is useful signal can be processed in your 16bit, 32bit or whatever advanced engine, this is the benefit I was keeping to bring it out.

So, to summarize what I think about this whole thing, at deep black, sensor noise level define tonal step. Move up the grey scale, having more bit ADC potentially have more tonal step which could be even smaller than deep step/noise level. No matter what is the definition it is/was, this is the point I was trying to make here.



Apr 24, 2013 at 05:31 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.3 #13 · Dynamic Range vs. Tonal Range


Since Alundeb is talking in linear (again, this NEEDS to be stated quite often), noise is stronger in bright areas than in dark areas. A very strange sounding concept, unless you keep linear and "normalized" values separated.

Let's say we "see" in Ev stops, gamma 2.0 (we don't, but just as an example). Then we'll "see" the difference between 20 and 40 as the same as the step between 200 and 400. Exactly one stop more.

The noise at 40 is sqrt(40) = 6.3
The noise at 400 is sqrt(400) = 20

Noise is stronger at 400 than at 40....

But at level 400 the noise is only 20/400 = 0.05 of the signal level
At level 40, those 6.3 is actually 6.3/40 = 0.16 of the signal level, almost three times more!

So we'll "see" the noise in the brighter area as lower in strength, even though it's a higher real number. Since noise needs to be thought of as relative to the current signal level.



Apr 24, 2013 at 05:35 PM
alundeb
Offline
• • • •
Upload & Sell: Off
p.3 #14 · Dynamic Range vs. Tonal Range


Yes, linear, thanks for pointing it out

zhangyue wrote:
Can I have some background information about noise source here: shot noise is from photons/sensor, so you mean more exposure will not only amplify signal, but also noise, so we are not change SN ratio here.


The noise is in the light even before it reaches the sensor. It is not about amplification, it is the statistical properties of random discrete particles.

http://en.wikipedia.org/wiki/Shot_noise



Apr 24, 2013 at 05:49 PM
RustyBug
Offline
• • • • • •
Upload & Sell: On
p.3 #15 · Dynamic Range vs. Tonal Range


So, does that mean for "shot noise", a longer exposure (equivalent exposure) will "even out" the randomness of the distribution somewhat ... but, it comes at the risk of incurring greater thermal noise at exaggerated shutter lengths?




Apr 24, 2013 at 06:14 PM
AhamB
Offline
• • • •
Upload & Sell: On
p.3 #16 · Dynamic Range vs. Tonal Range


zhangyue wrote:
But with limited bit at capture? I don't know how much benefit we get convert a 4 bits file to 16 bits and do PP there?


No... I'm talking about an ordinary 12 or 14 bits RAW file that you use in practice.



Apr 24, 2013 at 06:23 PM
alundeb
Offline
• • • •
Upload & Sell: Off
p.3 #17 · Dynamic Range vs. Tonal Range


RustyBug wrote:
So, does that mean for "shot noise", a longer exposure (equivalent exposure) will "even out" the randomness of the distribution somewhat ... but, it comes at the risk of incurring greater thermal noise at exaggerated shutter lengths?



The net effect of photon shot noise does not depend on the method you use to increase the exposure. If you use a longer exposure and smaller aperture, you get the same shot noise as if you use a shorter exposure and a larger aperture.



Apr 24, 2013 at 06:59 PM
zhangyue
Offline
• • • •
Upload & Sell: Off
p.3 #18 · Dynamic Range vs. Tonal Range


AhamB wrote:
No... I'm talking about an ordinary 12 or 14 bits RAW file that you use in practice.


For Low Dynamic range (4 stop) image, your 12 bit or 14 bit RAW have only 4 bit information. Most bits are not used. That is why (I think) we will be benefit having more bit raw file to start with.



Apr 24, 2013 at 07:03 PM
alundeb
Offline
• • • •
Upload & Sell: Off
p.3 #19 · Dynamic Range vs. Tonal Range


zhangyue wrote:
For Low Dynamic range (4 stop) image, your 12 bit or 14 bit RAW have only 4 bit information. Most bits are not used. That is why (I think) we will be benefit having more bit raw file to start with.


Even for a 4 stop DR scene (in the "Object" or "Subject" space) and a linear encoding, the number of bits used can be much higher than 4.

If the brightest and darkest value in a 14 bit raw file are say 8191 and 2047 (1 stop headroom for hgihlight clipping), 13 down to 11 bits are used to encode all values in between. You get about 6000 brightness steps.

If this image was taken with a camera with a good well capacity, say 65536 e- (photons) for saturation, the noise level measured in A/D units at the darkest value would be sqrt(65536 * 2047 / 16384) / (65536 / 16384) = 23
And the brightest value sqrt(65536 * 8191 / 16384) / (65536 / 16384) = 45

You get 23 to 45 steps to represent a change in brightness corresponding to the noise level.



Apr 24, 2013 at 07:51 PM
zhangyue
Offline
• • • •
Upload & Sell: Off
p.3 #20 · Dynamic Range vs. Tonal Range


alundeb wrote:
Even for a 4 stop DR scene (in the "Object" or "Subject" space) and a linear encoding, the number of bits used can be much higher than 4.

If the brightest and darkest value in a 14 bit raw file are say 8191 and 2047 (1 stop headroom for hgihlight clipping), 13 down to 11 bits are used to encode all values in between. You get about 6000 brightness steps.

If this image was taken with a camera with a good well capacity, say 65536 e- (photons) for saturation, the noise level measured in A/D units at the darkest value would be
...Show more

good trick but very nice catch I didn't go over your number here but you are right that for high key image, the step level's loss can be less. the lower key the image, the bigger impact with Less bit ADC. I was default use a low key image so that for 12 bit DR sensor with 12 bit ADC, if there is low key 4bit DR image, the step will be merely 8 step. instead of high key's 2048-256=1792 step.

Most people don't need go through these number but understand the consequence should be enough.

I was point this out early to Rusty's ETTR, but with unclear point that I shouldn't say remap to 8 bit tonal range.

zhangyue wrote:
ETTR has two merit here:
1. like you said, you are push the signal away from noise. (analog push)
2. it essentially maximize the ADC used in sensor: so for 4 bit DR image example aboe, you remap it to 8bit tonal range, so you can PP it as desired later. but your truncated noise is down to only 1/2^8 level.






Apr 24, 2013 at 08:52 PM
1       2      
3
       4       end




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

1       2      
3
       4       end
    
 

You are not logged in. Login or Register

Username   Password    Retrive password