High ISO vs Underexposure. What is best for noise?

[Guillermo Luijk]

Hi all, I have a question for those brilliant minds around here.

We all know that:
- The higher the ISO for a given exposure, the lower the SNR.
- The lower the exposure for a given ISO, the lower the SNR as well.

But if we are forced to use some exposure parameters (aperture and exposure time), what is recommended to obtain the best possible SNR (Signal to Noise Ratio), i.e. less visible noise, to push the ISO number? or to allow certain underexposure?

One could think both options would be the same in terms of visible noise, but this is not the case. I have done a test shooting the same scene with the same exposure parameters changing only the ISO, and the result was that the higher the ISO, the lower the visible noise, specially in the shadows.

Test: 3 shots at 70mm, 0.3s, f/8 using ISO 100, 200 and 400.

This is the scene:



And these are crops of normal, dark and very dark areas of the scene from top to bottom.
(development done with no noise reduction in DCRAW balancing exposures in PP):




I am sorry for the large file size due to noise. It is easy to see that the higher ISO is, the better SNR we achieve.

The fist conclusion I make of this is that it is better to use high ISO values than leaving our shot underexposed. Also that exposing to the right makes sense not only at ISO100 but at any ISO value, so ISO may be used to get ETTR when needed for being forced to use certain fixed exposure parameters (aperture and exposure time).

I wonder if it's the same for even higher ISO values, but I guess it is.

My question is: what explanation do you think this phenomemon has?
1. The camera applies some noise reduction in the signal amplification process when using high ISO (i.e. with large amplification values)?
2. There is some noise source that is added after amplification, so the higher the signal level is (thanks to ISO) the less that 'last minute' noise source will affect?

Regards

[ErikKaffehr]

Hi!

This is really an issue I wanted to investigate myself. I don't see any obvious answer. What camera are you using, BTW?

Two possible explanations would be that if the dynamic range of the AD-converter is less than the dynamic range of the sensor amplification befor AD-conversion would reduce noise. The other possible explanation I see would be that the sensor is doing some kind of noise reduction before outputting RAW.

It seems to me that no resolution is lost when going from 100 to 400 ISO. That may indicate that no excessive reduction in luminance noise is done in the sensor. I have no further ideas.

Best regards

Erik

Hi all, I have a question for those brilliant minds around here.

We all know that:
- The higher the ISO for a given exposure, the lower the SNR.
- The lower the exposure for a given ISO, the lower the SNR as well.

But if we are forced to use some exposure parameters (aperture and exposure time), what is recommended to obtain the best possible SNR (Signal to Noise Ratio), i.e. less visible noise? Push the ISO number? or to allow certain underexposure?

One could think both options would be the same in terms of visible noise, but this is not the case. I have done a test shooting the same scene with the same exposure parameters changing only the ISO, and the result was that the higher the ISO, the lower the visible noise, specially in the shadows.

Test: 3 shots at 70mm, 0.3s, f/8 using ISO 100, 200 and 400.

This is the scene:



And these are crops of normal, dark and very dark areas of the scene from top to bottom.
(development done with no noise reduction in DCRAW balancing exposures in PP):


I am sorry for the large file size due to noise. It is easy to see that the higher ISO is, the better SNR we achieve.

The fist conclusion I make of this is that it is better to use high ISO values than leaving our shot underexposed. Also that exposing to the right makes sense not only at ISO100 but at any ISO value, so ISO may be used to get ETTR when needed for being forced to use certain fixed exposure parameters (aperture and exposure time).

I wonder if it's the same for even higher ISO values, but I guess it is.

My question is: what explanation do you think this phenomemon has?
1. The camera applies some noise reduction in the signal amplification process when using high ISO (i.e. with large amplification values)?
2. There is some noise source that is added after amplification, so the higher the signal level is (thanks to ISO) the less that 'last minute' noise source will affect?

Regards
[a href=\"index.php?act=findpost&pid=153903\"][{POST_SNAPBACK}][/a]

[Slobodan Blagojevic]

It is rather simple, actually: underexposure will always produce more noise, especially in the shadows.

It has to do with the linearity with which the sensor captures light (try googling "linear RAW capture" or read this: http://www.adobe.com/digitalimag/pdfs/linear_gamma.pdf). In other words, and very simplified, the vast majority of information is captured in the brightest steps, and conversely, the least amount of tonal information is available in the darkest areas.

[Guillermo Luijk]

Hi slobodan, I think we have already gone a bit beyond your point. It is well known that overexposed images have greater signal to noise ratio, but this is clear when you achieve overexposure through real exposure camera parameters (aperture, exposure time).

ISO is some kind of artificial electronic amplification of the captured signal, and we wonder what's the reason inside the sensor electronics to improve visible noise when making use of ISO to achieve the overexposure. It's not that clear.

BTW Erik I have a 350D (CMOS).

Regards.

[ErikKaffehr]

Hi!

I checked this site:
http://www.clarkvision.com/imagedetail/dig...rmance.summary/

It seem that the 350D has a well capacity of 43000 electrons an a read noise of 3.7 leading to a dynamic range of 13.5 steps, which is a bit more than the DR for a 12bit converter which is 12 steps. So in this case amplifying the signal before conversion may actually reduce noise.

That said I'm still surprised!

Regarding the experiment it seems to me that you have eliminated all probable "cheat" factors by using a standard RAW-converter instead of a commercial one and keeping exposure fixed.

Best regards

Erik

Hi slobodan, I think we have already gone a bit beyond your point. It is well known that overexposed images have greater signal to noise ratio, but this is clear when you achieve overexposure through real exposure camera parameters (aperture, exposure time).

ISO is some kind of artificial electronic amplification of the captured signal, and we wonder what's the reason inside the sensor electronics to improve visible noise when making use of ISO to achieve the overexposure. It's not that clear.

BTW Erik I have a 350D (CMOS).

Regards.
[a href=\"index.php?act=findpost&pid=153960\"][{POST_SNAPBACK}][/a]

[Guillermo Luijk]

Hi!

I checked this site:
http://www.clarkvision.com/imagedetail/dig...rmance.summary/

It seem that the 350D has a well capacity of 43000 electrons an a read noise of 3.7 leading to a dynamic range of 13.5 steps, which is a bit more than the DR for a 12bit converter which is 12 steps. So in this case amplifying the signal before conversion may actually reduce noise.

That said I'm still surprised!

Regarding the experiment it seems to me that you have eliminated all probable "cheat" factors by using a standard RAW-converter instead of a commercial one and keeping exposure fixed.

Best regards

Erik
[a href=\"index.php?act=findpost&pid=153966\"][{POST_SNAPBACK}][/a]

I got to the same website this afternoon. However I had to stop as I was having a terrible headache hehe. Will give it a second try.

Yes I used DCRAW which by default performs no noise reduction and scaled all images by the same white balance factors. Next I balanced the ISO100 image to match the ISO400 with a +2EV curve, and the ISO200 image with a +1EV curve.

The crops are 50% the real size, but resized using nearest neighbour (no interpolation) so SNR remains statistically the same.

Could you please explain the mechanism why SNR is improved thanks to the DR compression? I am an engineer so you can talk in electronic terms.

[ErikKaffehr]

Hi,

Sorry, I have no good explanation. My reasoning is that noise floor is below the threshold of the sensor, so some of the noise may just originate from the LSB going either one or zero. It does not really see to be a reasonable explanation to me, but I felt I could mention it anyway.

Best regards

Erik

I got to the same website this afternoon. However I had to stop as I was having a terrible headache hehe. Will give it a second try.

Yes I used DCRAW which by default performs no noise reduction and scaled all images by the same white balance factors. Next I balanced the ISO100 image to match the ISO400 with a +2EV curve, and the ISO200 image with a +1EV curve.

The crops are 50% the real size, but resized using nearest neighbour (no interpolation) so the statistical SNR is safely kept.

Could you please explain the mechanism why SNR is improved thanks to the DR compression? I am an engineer so you can talk in electronic terms.
[a href=\"index.php?act=findpost&pid=153974\"][{POST_SNAPBACK}][/a]

[Guillermo Luijk]

Hi,

Sorry, I have no good explanation. My reasoning is that noise floor is below the threshold of the sensor, so some of the noise may just originate from the LSB going either one or zero. It does not really see to be a reasonable explanation to me, but I felt I could mention it anyway.

Best regards

Erik
[{POST_SNAPBACK}][/a]

Makes sense. However I have just received a quite convincent answer. the reason for the improvement is my second point: a noise source being added after signal amplification and before A/D (probably in the A/D itself):

...that noise source is probably in the ADC, although other analog circuitry could contribute as well. (I'll call it "ADC noise" regardless.)

Most cameras do not apply "lossy" noise reduction algorithms to their RAW data.

> I wonder if it's the same for even higher ISO values, but I guess it is.

Actually it isn't. Increasing the ISO amplifies both the signal and the sensor read noise (pre-amplification) relative to the ADC noise. When sensor read noise becomes the dominant contribution to shadow noise, there is no further advantage for higher ISO at the same absolute exposure.

Roger Clark has some interesting information about this stuff:
[a href=\"http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html]http://www.clarkvision.com/imagedetail/eva...-1d2/index.html[/url]

--
Alan Martin


In simple terms I think this is the idea: the input to the ADC is:

OUTPUT = (S_in + N_read) * ISO + N_adc

SNR= (S_in*ISO) / (N_read*ISO + N_adc) which becomes higher the higher ISO is.

But if we push ISO a lot amplified N_read becomes dominant over N_adc:

N_read*ISO >> N_adc -> SNR ~ S_in / N_read which cannot be improved any more.

[allan67]

Hello,

There was a discussion about this somewhere on LL forum (can't find the thread now). The conclusion was as you put it - it's better to have analog amplification in camera (using higher ISO) than a digital curve applied in post-production.

If I remember correctly, it was also stated, that it's better to use whole multiples when going to higher ISOs - for base 100 you should use 200, 400, 800, 1600 and not intermediate values (like 320, 640 etc). That is apparently due to implementation of this feature in cameras - the circuitry implementing whole multiples seems to be better than for intermediate values.
Don't know if this is really true, I see MR uses a lot of intermediate ISO values when shooting, so this might not be an issue in real-world situations.

Allan

[bjanes]

Hi all, I have a question for those brilliant minds around here.

We all know that:
- The higher the ISO for a given exposure, the lower the SNR.
- The lower the exposure for a given ISO, the lower the SNR as well.

But if we are forced to use some exposure parameters (aperture and exposure time), what is recommended to obtain the best possible SNR (Signal to Noise Ratio), i.e. less visible noise, to push the ISO number? or to allow certain underexposure?

I wonder if it's the same for even higher ISO values, but I guess it is.
[{POST_SNAPBACK}][/a]

Hi Guillermo,

Your test confirms what [a href=\"http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html]Roger Clark[/url] has predicted.

According to his analysis, signal:noise in the highlights is determined by photon counting statistics, but read noise predominates in the deep shadows. Since the exposure (integration time) was the same for both ISOs, the noise should be the same in the highlights, since the number of photons collects was the same. However, at high ISO the read noise decreases and shadow noise is lessened. From a visual examination of your results, it looks to me that the noise differences in the two ISOs are mainly in the shadow areas as predicted.

In another essay, Roger discusses unity gain. He says that above unity gain, there will be no further improvement by increasing the ISO setting. However, he notes that you lose 1 stop of dynamic range for each doubling of ISO.

For the Canon 1D MII, he suggests that ISO 1600 is optimum and has the lowest read noise, but that ISO 800 is nearly as good and has almost double the dynamic range.

Bill

[ePaperPress]

Here is a test I did on the Canon G9. ISO 1600 shots were properly exposed, while ISO 800 shots were under-exposed by one stop. Raw images were processed per the Photoshop preset described at

http://forums.dpreview.com/forums/read.asp...essage=25612433

Exposure was increased in the ISO 800 Raw image by +1. JPEG images were postprocessed using Noise Ninja with default settings. The ISO 800 image was modified with Curves to correct exposure. Images were downsized 50% and cropped.

JPEG 1600:


JPEG 800:


Raw 1600:


Raw 800:


Conclusion:
ISO 800 images fared better than ISO 1600 images.
Raw images fared better than JPEG images.

[Guillermo Luijk]

Hi ePaper, the result is as you say, lower ISO = lower noise. But the way to get to the final images seems too confusing and incorrect to me. Let me explain:

Basic:
  Clarity: 25
Detail:
  Amount: 60
  Radius: .8
  Detail: 30
  Masking: 0
  Luminance: 20 (80 for Noise Reduction)
  Color: 100

What is Detail for instance? Noise reduction? how do you know if noise reduction or that detail parameter are applied the same way to 2 differently exposed and ISO images?

I don't think this is the right way to make a noise comparision. I suggest to develop the RAW files with a totally neutral RAW developer (DCRAW is a fine tool for this), just adjust exposure, and show crops using nearest neighbour scaling (not interpolation as you used, which improves SNR).

If you make the RAW files available I would be glad to do all these steps.

Thanks

[ePaperPress]

Hi ePaper, the result is as you say, lower ISO = lower noise. But the way to get to the final images seems too confusing and incorrect to me. Let me explain:

Basic:
  Clarity: 25
Detail:
  Amount: 60
  Radius: .8
  Detail: 30
  Masking: 0
  Luminance: 20 (80 for Noise Reduction)
  Color: 100

What is Detail for instance? Noise reduction? how do you know if noise reduction or that detail parameter are applied the same way to 2 differently exposed and ISO images?

[{POST_SNAPBACK}][/a]

You'll find details for Photoshop's settings here:

[a href=\"http://photoshopnews.com/2007/05/31/about-camera-raw-41/]http://photoshopnews.com/2007/05/31/about-camera-raw-41/[/url]

For me it doesn't matter what the results are without sharpening and noise reduction applied since that's my workflow. I can understand how others may disagree. You may contact me at the Feedback link at http://epaperpress.com/whoami and I'll link you to a download that contains the raw files. Be advised that they've been converted to DNG format

[Guillermo Luijk]

You'll find details for Photoshop's settings here:

http://photoshopnews.com/2007/05/31/about-camera-raw-41/

For me it doesn't matter what the results are without sharpening and noise reduction applied since that's my workflow. I can understand how others may disagree. You may contact me at the Feedback link at http://epaperpress.com/whoami and I'll link you to a download that contains the raw files. Be advised that they've been converted to DNG format
[a href=\"index.php?act=findpost&pid=154180\"][{POST_SNAPBACK}][/a]

But if through your workflow you get a poorer image from the highest ISO shot, which in neutral circumstances could be less noisy, it means you could be doing a non-optimum workflow and that you can get better results from your images.

I will go to your thread... OK I sent a feedback message. Are you the author of PTLens? wow, great software indeed!

PS: NP for DNG, DCRAW perfectly supports that format.

[ePaperPress]

For me it doesn't matter what the results are without sharpening and noise reduction applied since that's my workflow. [a href=\"index.php?act=findpost&pid=154180\"][{POST_SNAPBACK}][/a]

Well, you got me interested. So I set all sliders that had an effect on sharpness and noise reduction to zero in ACR. Here are a few 100% crops. Note that Exposure, for the 800 image, was set at +1.

ISO 800:


ISO 1600:


It seems there is a lot of chroma noise in the 1600 image and colors are slightly different even though they had the same white balance. This is done in ACR and there may be other factors at work here that I'm not aware of.

[Guillermo Luijk]

Yeah, definitively your ISO800 is better than your ISO1600. I would love to try them on DCRAW but looking at this I don't think we will reach a different conclusion. This can be very interesting as it means the behaviour of ISO is different according to vendor.

What camera model do you have? CMOS or CCD? Mine is a 350D (CMOS ).

These are my RAW files in case anyone is interested in giving them a try:

http://www.guillermoluijk.com/download/iso100.cr2

http://www.guillermoluijk.com/download/iso200.cr2

http://www.guillermoluijk.com/download/iso300.cr2

[ePaperPress]

What camera model do you have? CMOS or CCD? Mine is a 350D (CMOS ).
[a href=\"index.php?act=findpost&pid=154190\"][{POST_SNAPBACK}][/a]

The images posted here are for the Canon G9 that has a small 1/1.7 CCD. As we say, we're trying to squeeze blood from a turnip here!

[Guillermo Luijk]

The images posted here are for the Canon G9 that has a small 1/1.7 CCD. As we say, we're trying to squeeze blood from a turnip here!
[a href=\"index.php?act=findpost&pid=154195\"][{POST_SNAPBACK}][/a]

Well here is my test over your images. Although a bit more neutral it still shows that your ISO800 is better than your ISO1600 image, what makes me think the behaviour regarding ISO is in close relation to the sensor technology (I don't think this is a matter of internal PP since I believe cameras don't perform any NR over the RAW data).

As in your test, there is quite a lot more colour noise in the ISO1600 imagen than in the ISO800 image:

 

[allan67]

Anyone can explain this?   
[a href=\"index.php?act=findpost&pid=154203\"][{POST_SNAPBACK}][/a]

Hello,

How is ISO 1600 implemented in G9? Is it true analog gain boost or just a trick, like ISO 3200 on 20D?
If it's the latter, this might explain the outcome. I have tried the experiment like this myself with my 30D and the jump in SNR between 1600 and 3200 is MUCH greater than between 800 and 1600.

Allan

[ePaperPress]

How is ISO 1600 implemented in G9? Is it true analog gain boost or just a trick, like ISO 3200 on 20D?
[a href=\"index.php?act=findpost&pid=154324\"][{POST_SNAPBACK}][/a]

It's difficult to say without knowledge of the G9 circuitry. The G9 has an ISO knob that goes from 80 to 1600. There's a special scene setting that allows you to shoot at 3200 at reduced resolution.

At any rate, when I need 1600 I just dial in -1 on exposure compensation and shoot at 800 for better results.