Noise: ISO push vs. PP push

[Stosh]

I've seen it mentioned here several times.
A properly exposed high ISO image can have less noise than an improperly (dimly) exposed low ISO image.
Why would this be? I'm not saying it's wrong, just trying to understand why. I'd like to look at as fair of an example as I can think of. Let's start by properly exposing a scene at... say ISO 3200. Then let's keep aperture and shutter speed identical and reduce the ISO 2 stops to 800. What should happen here?

The first thing that comes to mind is my experience with Canon DSLRs and Canon's PP software. In their Digital Photo Professional, you can "push" your exposure by 2 stops and it "magically" brings out those 2 stops of previously invisible data. So the data is there even though you couldn't see it before pushing. Yes, it's definitely noisier than the same image with no push, but is it noisier than the properly exposed image at ISO 3200?

Another observation is that Digital Photo Professional's noise reduction settings. If you take an exposure at ISO 800 by default it will use very little noise reduction. At ISO 3200 it will use a significant amount of noise reduction. So the comparison isn't fair until you bump up the noise reduction on the pushed image to match that of the ISO 3200 image.

The way I understand digital ISO is that the signal from the sensor is amplified higher (multiplied) at higher ISOs. In post processing to achieve a brighter image you do the same thing in the digital realm (multiply the values). Assuming I'm right about how digital ISO works, why would the analog gain from the sensor be less noisy than the digital gain in post processing?

 

[Big Mike]

The more 'underexposed' your image is, the lower your signal to noise ratio will be...because there is just less information in the darker ranges of the digital photo.

Expose Right

 

[Stosh]

The more 'underexposed' your image is, the lower your signal to noise ratio will be...because there is just less information in the darker ranges of the digital photo.

Expose Right

I understand that, but how would increasing the gain from the sensor's signal do any better with the signal to noise ratio? When you amplify a signal, you amplify both signal AND noise therefore not effectively gaining anything.

 

[JG_Coleman]

Because of the way a digital sensor records image data, there is exponentially less information in progressively darker portions of a photograph.

Let's say you've got an underexposed digital image on your computer and there's a huge, dark shadow that you want to brighten up. The problem is that, even though the RAW file does indeed have some data stored for details within this shadow, it is extremely minimal. When you brighten up the shadow in the RAW file, you're basically amplifying the brightness of a nearly pure black area of the digital photograph.

Let me try to clarify this... Consider the opposite end of the spectrum. If you have a digital image with blown-out highlights... those areas of the photo contain little or no information that can be recovered. They've been recorded by the camera as pure white (or nearly pure white) blotches that lack usable detail. The same goes for shadows, really. If any area is recorded as pure black (or nearly pure black) there's going to be very little saved data in that area of the photo to recover. Thus, when you brighten up a dark shadow, you will recover some detail... but it will be grainy, noisy, and lack acceptable color fidelity.

Now, what if we use a high ISO? From the start, this seems to put you at a severe disadvantage, since the higher ISO will produce a noisier image by default. While that's true, there's still an important difference between the noise produced by a high ISO shot and the noise produced by boosting exposure in post.

The difference is in the source of the data. You'll be able to pull better detail from shadow areas using a higher ISO, because that extra exposure achieved will be able to call upon the actual light coming from the subject. Even though high ISO shots are noisy, the shadow areas will still have a better signal-to-noise ratio than if you simply underexposed the shot and brightened it up in post. Adding exposure in post, if you try to go too far, can be equivalent to trying to "squeeze blood from a stone". The digital image can only hold so much data... there IS a point at which areas which are mostly black just don't have any more usable data to reveal. In contrast, the world itself possesses nearly infinite detail in the form of reflected light. When you're using a higher ISO, the subject you're trying to record is right in front of the camera, ready to yield all the details you desire if you only expose the shot thoroughly enough.

 

[Jay Hsiang Studio]

helpful post. thanks!

georgia from jay hsiang studio

Jay Hsiang Studio
Los Angeles Wedding Photography
Engagement Photography
Destination Wedding Photography

 

[Stosh]

Because of the way a digital sensor records image data, there is exponentially less information in progressively darker portions of a photograph.

Yup, totally agree. Because of the linear nature of the sensor, each darker stop has half as much info as the previous brighter stop.

Let's say you've got an underexposed digital image on your computer and there's a huge, dark shadow that you want to brighten up. The problem is that, even though the RAW file does indeed have some data stored for details within this shadow, it is extremely minimal. When you brighten up the shadow in the RAW file, you're basically amplifying the brightness of a nearly pure black area of the digital photograph.

Let me try to clarify this... Consider the opposite end of the spectrum. If you have a digital image with blown-out highlights... those areas of the photo contain little or no information that can be recovered. They've been recorded by the camera as pure white (or nearly pure white) blotches that lack usable detail. The same goes for shadows, really. If any area is recorded as pure black (or nearly pure black) there's going to be very little saved data in that area of the photo to recover. Thus, when you brighten up a dark shadow, you will recover some detail... but it will be grainy, noisy, and lack acceptable color fidelity.

Yup, again we agree. What you call grain, I'll call posterization. Yes, pushing dark areas definitely results in more noise than brighter areas. Great, we agree here.

Now, what if we use a high ISO? From the start, this seems to put you at a severe disadvantage, since the higher ISO will produce a noisier image by default. While that's true, there's still an important difference between the noise produced by a high ISO shot and the noise produced by boosting exposure in post.

I'm all ears.

The difference is in the source of the data. You'll be able to pull better detail from shadow areas using a higher ISO, because that extra exposure achieved will be able to call upon the actual light coming from the subject.

As opposed to........ light coming from somewhere else? I don't understand what you're saying here.

Even though high ISO shots are noisy, the shadow areas will still have a better signal-to-noise ratio than if you simply underexposed the shot and brightened it up in post.

You've just re-worded my question into a statement. How does that answer my question?

Adding exposure in post, if you try to go too far, can be equivalent to trying to "squeeze blood from a stone". The digital image can only hold so much data... there IS a point at which areas which are mostly black just don't have any more usable data to reveal.

Yup, totally agree. There is definitely a point in which you can not squeeze any more data out of the black areas. But how does this answer my question? My example uses 2 identical exposures. One pushes 2 stops through ISO and one pushes 2 stops through PP. I don't think you're telling me 2 stops in PP is like getting blood from a stone. Certainly 6 stops is, but I realize that's way above and beyond the capability of good PP. A 6-stop boost from ISO 100 is ISO 6400 which is noisy, yes, but at least you have SOMETHING. If you push an image in PP by 6 stops you won't have very much. I never suggested that PP pushing could match ISO pushing in extreme situations. I tried to make the experiment fair because Canon seems to record at least 2 stops of shadow information that you can't even see until you use RAW processing to bring it out. You just can't bring out any "hidden" shadow info from a JPG. It doesn't exist.

In contrast, the world itself possesses nearly infinite detail in the form of reflected light. When you're using a higher ISO, the subject you're trying to record is right in front of the camera, ready to yield all the details you desire if you only expose the shot thoroughly enough.

This was a nice summary, but again, I can't find an answer to my original question. Please help me if I just didn't see it.

 

[JG_Coleman]

Okay... the problem here is really just that, on the fly, it was a difficult concept to convey. What I'll try to do is explain it a different way, with a specific example first.

To illustrate the concept, it's necessary to use a somewhat exaggerated example. Bear in mind that this example is put forth just for the sake of illustrating the concept.

Let's say your in this scenario.

-You're taking a landscape photo on a beach at night illuminated dimly by the moon.
-You're working with an aperture of f/5.6 on a kit lens.
-You're working with a specific shutter speed... let's say 1/100 sec.

Now, imagine you take a photograph at ISO 100 with these settings. You'll pretty much get a pitch black frame with almost no information at all. If you try to push the exposure through software, you might get hazy shapes of the beach and the horizon... but it could barely be called a photograph. The data to draw the image out just isn't there.

Now, let's say that without changing the aperture or the shutter speed, you take the exact same shot at ISO 12,800. It will be a seriously grainy photograph... pure crap as far as a good photo goes... but you'll actually be able to see the scene rather well in comparison to the first shot.

Now, as I said earlier, this is a severely exaggerated scenario. But it basically takes the issue of "ISO push vs. PP Exposure Push" to its most extreme in order to demonstrate the inherent weakness in pushing exposure in post.

Quote: Originally Posted by JG_Coleman
The difference is in the source of the data. You'll be able to pull better detail from shadow areas using a higher ISO, because that extra exposure achieved will be able to call upon the actual light coming from the subject.


As opposed to........ light coming from somewhere else?

As opposed to having a photograph that is already exposed and cannot be further exposed to any light whatsoever. That's the whole reason that the PP exposure push doesn't really stack up to ISO push.

Using our above example, why is it that the ISO 100 shot can't even come close to the ISO 12,800 shot?

With the ISO 100 shot, the shutter opened up for 1/100 sec and the sensor basically saw absolutely nothing... and that's what it recorded... nothing. That's all it has... nothing can be drawn out, because nothing was recorded.

With the ISO 12,800 shot, the shutter opened up for 1/100 sec and the sensor saw, through plenty of color noise, a beach scene complete with sand and water and a few dunes.

The way I understand digital ISO is that the signal from the sensor is amplified higher (multiplied) at higher ISOs. In post processing to achieve a brighter image you do the same thing in the digital realm (multiply the values).

This is where you get confused...

You're thinking of using high ISOs as "boosting" the brightness of the image. In fact, that is not what's happening. High ISOs don't "boost" anything except for the sensitivity of the sensor. It's not as if the camera is just applying its own post-processing to make the image brighter. The sensor itself is literally made more sensitive to light, and therefore exposes faster than it would otherwise. It has the same basic effect as if you left your ISO low and instead left the shutter open longer. High ISO doesn't multiply data... it literally collects more data. It does this at the expense of generating color noise... but it is nonetheless collecting more data.

Using post-processing to increase exposure, in contrast, is just a boost. You're just brightening up... multiplying... the minimal data available.

If you can expose your photograph enough at the time of capture, either by lengthening shutter speed or increasing ISO, you are gathering MORE source data (in the form of reflected light). If you increase exposure in post, you are not gathering any more source data... you're just brightening up what minimal data you already have.

Admittedly, this is a difficult thing to explain. I feel like there's an ideal way to succinctly express this concept that I'm just not able to draw out of myself, which is why I'm left to make meandering explanations.

Maybe this explains it a bit better?

 

[JG_Coleman]

So, given a situation where you can shoot one of two photos:

Photo A - shoot two stops underexposed with low ISO (then boost exposure in post)

Photo B - shoot properly exposed at a higher ISO

Quite frankly, at only two stops underexposure, there is probably not an enormous difference between the two. I would still say, based on my previous points, that Photo B will be the better photograph overall. That the RAW file records "at least 2 stops of shadow information" is something of a half-truth. True, details can be pulled from shadows that are two stops under-exposed... but the information in those under-exposed areas is of markedly lower quality than anything else in the image that is properly exposed. And the closer an area is to being a full two stops under-exposed, the worse it gets.

Brighter areas of a photograph hold exponentially more data than darker areas, as you've agreed. Using higher ISO to properly expose a scene simply captures much more data... it's really just that simple. That high ISOs generate color noise is a point not to be ignored... but the signal-to-noise ratio is still higher than if you start out with minimal signal to begin with in an under-exposed shot.

I understand that I'm referring to extreme circumstances, and that you're talking about a much less extreme PP exposure push. The fact of the matter, though, is that the principle found in extreme examples carries right over to less extreme examples. The difference between Photo A and Photo B won't be as pronounced as the ISO 100 vs ISO 12800 shot I used as an example, but the same principle carries right up to the difference between ISO 200 and ISO 400.

Both areas are going to show noise in their own way... but the properly exposed shot will have much more signal mixed in with that noise.

 

[Garbz]

More signal is right. That's the problem here. The amplification change in the ISO is not just a software processed difference of the signal. It changes the way the signal is amplified and converted to yield better signal to noise ratio than simply bumping the brightness of the data already captured.

I see a lot of discussion here, but not a lot of testing. Let's fix that
Here is our base image, shot at ISO100.

Here is in order the image with shutterspeed reduced 2 stops:
Firstly shot at ISO400.
Secondly shot at ISO100, with the exposure slider at +2 fill light cranked up, saturation reduced and contrast reduced.
Finally shot at ISO100, with JUST the exposure slider at +2 (note how the tone is totally wrong. Why Adobe do you do this? Why?)

Here is in order the image with shutterspeed reduced 4 stops:
Firstly shot at ISO1600.
Secondly shot at ISO100, with the exposure slider at +4 fill light cranked up, saturation reduced and contrast reduced.
Finally shot at ISO100, with JUST the exposure slider at +4 (note even more how screwed the basic exposure slider is compared to a real change in exposure)

Please ignore motion blur. Desk doesn't a good tripod make

 

[Stosh]

With the ISO 100 shot, the shutter opened up for 1/100 sec and the sensor basically saw absolutely nothing... and that's what it recorded... nothing. That's all it has... nothing can be drawn out, because nothing was recorded.

With the ISO 12,800 shot, the shutter opened up for 1/100 sec and the sensor saw, through plenty of color noise, a beach scene complete with sand and water and a few dunes.

OK, we're making progress because we found something we disagree on. How did these 2 identical sensors with identical shutter speeds see 2 different things? Back in film days I would agree with you 100%. The film was vastly different in sensitivity to light depending on which film you loaded behind the shutter. But today you don't change the sensor. It sees what it sees and that's that. The way I understand how ISO works, the gain is turned up on the sensor to make it more sensitive. Gain is another way of saying amplifier which is another way of saying multiplying which is another way of saying post processing (although it's post processing that the camera does on its own and we have no control over it).

This is where you get confused...

You're thinking of using high ISOs as "boosting" the brightness of the image. In fact, that is not what's happening. High ISOs don't "boost" anything except for the sensitivity of the sensor. It's not as if the camera is just applying its own post-processing to make the image brighter. The sensor itself is literally made more sensitive to light, and therefore exposes faster than it would otherwise. It has the same basic effect as if you left your ISO low and instead left the shutter open longer. High ISO doesn't multiply data... it literally collects more data. It does this at the expense of generating color noise... but it is nonetheless collecting more data.

It seems this is the meat of our discussion. Again, I don't agree with you here and would love to find out which side is more accurate. Is there anywhere to research this further?

That the RAW file records "at least 2 stops of shadow information" is something of a half-truth.

Maybe we need to explore this a little further too. The whole reason RAW is better than JPG is because it has a wider dynamic range (similar to how a negative held more info than a print).

If we express dynamic range in bit depth, it's easier to continue the discussion. A normal JPG that you see on a monitor or send to a printer is 8 bits per color channel. Off the top of my head, I don't know what the "standard" is for today's DSLRs, but to be safe, let's just say it's 10 bits of dynamic range. Let's say for your final picture you have only 2 options in post processing (we're using budget software lol). Option A is to use the lower 8 bits and option B is to the upper 8 bits. In both cases your final picture is using real world actual data that the sensor recorded. To use your words again, we're not trying to squeeze blood from a rock. Admittedly, the info in the bottom 8 bits will have a slightly lower SNR, but I think both would be excellent quality photos. Can I also say that photo B is "pushed" 2 stops beyond photo A because we're using a 2 bit shift? Can I describe this as "lossless pushing?"

As I've already said, I agree that the upper levels hold much more information than the lower levels, but I think that concept is greatly exaggerated and I think my case above shows why. Since a JPG only shows 8 bits per channel, you're not losing anything at all by using the lower 8 bits vs. the upper 8 bits. Sure, you're truncating the top 2 bits, but if there was nothing in those fabulously data-rich upper 2 bits, you're not losing a thing. You still end up with a full 8 bits of actual real-world data.

I contend that you lose exactly the same amount of data by dropping the upper 2 bits or the lower 2 bits. Again, I agree the lower 2 bits will have slightly more noise.

I understand that I'm referring to extreme circumstances, and that you're talking about a much less extreme PP exposure push. The fact of the matter, though, is that the principle found in extreme examples carries right over to less extreme examples.

I'm all for that line of thinking. I use extremes to illustrate subtle differences all the time. And in your extreme examples, I follow you 100% and agree with you 100%. But it's likely I didn't state my original thought well enough. I don't want to compare ISO push vs. PP push in situations where no info was recorded (which would be beyond however wide the dynamic range of the RAW output is). I want to compare true, recorded data to true, recoded data using 2 different methods. And just to be clear, here are the 2 different methods:

Method A: properly expose a scene in a higher ISO to use the upper 8 bits.
Method B: use a lower ISO to underexpose the same image only enough so that data is still recorded so that when you push the exposure in PP using the RAW file, you're not inventing data, you're using what's already there.

I think there are 2 keys to my viewpoint. #1 is that I think sensor gain (ISO) is just multiplying the same internal data. #2 using the lower 8 bits of a RAW image does not make a significant difference in using the upper 8 bits.

After re-reading my post, I think I may see where I have 1 concept not quite right. When I talk about upper 8 bits vs lower 8 bits, I'm thinking in the linear world which is definitely how the sensor records the data. However, the camera applies curves to this linear data to make the final picture look more "real world" the way our eyes perceive the scene. This conversion from linear to real-world curve may have something to do with my confusion. Can you help with that?

 

[Stosh]

I knew I could get you to post, Garbz!
Thanks for the examples. I like your use of subjects lol. You beat me to the hands-on portion, but I'm still interested in the why.

What about my noise reduction discussion? Doesn't the camera automatically apply more noise reduction when you use a higher ISO? And did you apply that same increased noise reduction to the PP pushed image to make it a little more fair?

 

[Helen B]

If we express dynamic range in bit depth, it's easier to continue the discussion. A normal JPG that you see on a [COLOR=#990000 ! important][COLOR=#990000 ! important]monitor[/COLOR][/COLOR] or send to a printer is 8 bits per color channel. Off the top of my head, I don't know what the "standard" is for today's DSLRs, but to be safe, let's just say it's 10 bits of dynamic range. Let's say for your final picture you have only 2 options in post processing (we're using budget software lol). Option A is to use the lower 8 bits and option B is to the upper 8 bits. In both cases your final picture is using real world actual data that the sensor recorded. To use your words again, we're not trying to squeeze blood from a rock. Admittedly, the info in the bottom 8 bits will have a slightly lower SNR, but I think both would be excellent quality photos. Can I also say that photo B is "pushed" 2 stops beyond photo A because we're using a 2 bit shift? Can I describe this as "lossless pushing?"

After re-reading my post, I think I may see where I have 1 concept not quite right. When I talk about upper 8 bits vs lower 8 bits, I'm thinking in the linear world which is definitely how the sensor records the data. However, the camera applies curves to this linear data to make the final picture look more "real world" the way our eyes perceive the scene. This conversion from linear to real-world curve may have something to do with my confusion. Can you help with that?

I think that understanding the linear (or near-linear) nature of the output from the ADC is the key to this discussion. Sending the ADC a full-range analogue signal is clearly a good thing.

Yes, a JPEG has a tone curve applied, it is not a simple truncation of the linear data. There is, therefore, no fixed relationship between the scene brightness range that a JPEG represents and the 8-bit nature of a JPEG.

Best,
Helen

 

[JG_Coleman]

Well... I'm running out of ways to explain it, really. But upon further contemplation, I think I've hit on the best way to explain it. Maybe this still won't work for you, but it's all I can really think of.

I've been somewhat simplifying the case of high ISO, but maybe being more technically specific information will help. Indeed, when you change the ISO to a higher value... you aren't technically increasing the sensor's sensitivity. You are right in that respect... you can't actually make the sensor more sensitive. Indeed, the signal is, in the strictest sense, just being amplified.

But, higher ISOs are amplifying the analog signal received by the sensor. This amplification is occurring before the data is sent to the A/D converter to be recorded as digital information.

So, when you use higher ISO to fully expose those extra two stops, you're drawing the extra two stops of information directly from the analog signal received by the sensor.

In contrast, when you underexpose by two stops with lower ISO and increase the exposure in post, you're amplifying digital information.

This is really the crux of the matter. The analog signal carries more image data than the digital information in a photograph that has already been captured.

 

[Stosh]

More signal is right. That's the problem here. The amplification change in the ISO is not just a software processed difference of the signal. It changes the way the signal is amplified and converted to yield better signal to noise ratio than simply bumping the brightness of the data already captured.

I see a lot of discussion here, but not a lot of testing. Let's fix that
Here is our base image, shot at ISO100.

Here is in order the image with shutterspeed reduced 2 stops:
Firstly shot at ISO400.
Secondly shot at ISO100, with the exposure slider at +2 fill light cranked up, saturation reduced and contrast reduced.
Finally shot at ISO100, with JUST the exposure slider at +2 (note how the tone is totally wrong. Why Adobe do you do this? Why?)

Here is in order the image with shutterspeed reduced 4 stops:
Firstly shot at ISO1600.
Secondly shot at ISO100, with the exposure slider at +4 fill light cranked up, saturation reduced and contrast reduced.
Finally shot at ISO100, with JUST the exposure slider at +4 (note even more how screwed the basic exposure slider is compared to a real change in exposure)

Please ignore motion blur. Desk doesn't a good tripod make

Now that I had a few minutes to look at this, I noticed something interesting. Of course the first thing I looked at was the dark area in the top center. Concerning noise in dark areas, your pictures certainly show that ISO 400 did better than ISO 100 pushed 2 stops. But now look down at the bottom book. Look at the word "Mathematics". Notice how much sharper it is in the ISO 100 shots? Both methods of pushing ISO 100 were sharper than the ISO 400 shot. My guess is back to what I said about noise reduction. The camera (or some PP) applied more noise reduction to the ISO 400 shot automatically. The noise reduction was not applied to the ISO 100 shots. So although the noise reduction certainly did its job of reducing noise in the dark areas, it significantly wounded the brighter part of the image by reducing sharpness.

So back to what I suggested before - if you were to somehow find the amount of noise reduction that got applied to the ISO 400 shot, then also applied it to the pushed ISO 100 shot, would it look similar?

 

[JG_Coleman]

#1 is that I think sensor gain (ISO) is just multiplying the same internal data.

I'm telling you, Stosh... you can look at sample shots all day long... but the conceptual under-pinnings are all you really need to know to understand the answer to your question. Your above assumption is incorrect.

When you use a high ISO, you're amplifying the original, analog signal produced by the camera's sensor. When you compensate for under-exposure in post, you're amplifying a weak digital representation of the original signal.

Thus, it's most definitely not "multiplying the same internal data". The two data sources are much different. One is the original signal produced by light striking your sensor, the other is an under-exposed, low-data, digital record of that signal.

Noise reduction may have it's various implications, as well. But the fact still remains... you get more image data by boosting the original analog signal than by boosting the low-data digital representation produced by under-exposure.