Does noise average out with exposure time?

[TonyW]

I have been wondering for a while whether noise in an image would be less with a longer exposure time for the same total amount of light.

In a recent thread, if I understand correctly, Colin Southern suggested that noise could be reduced with a large number of shots of the same scene stacked in PP.

I did some experiments. I photographed a blank white wall at f/2.8 SS1/1000, f/5.6 SS1/250 and f/11, SS1/60, all at ISO 1000, with my 100mm macro lens. The total exposure came out at about the same in each case as expected. There was no discernible difference in the amount of noise in the three cases.

This indicates to me that the noise does not get averaged out with time and makes me wonder what causes noise. I would have expected thermal/electronic noise to average out. Is it variations in the pixel characteristics?

Can anyone enlighten me on these questions?

[Kodiak]

I have been wondering for a while whether noise in an image would be less with a longer exposure time for the same total amount of light.

The presence of noise in a photograph is software specific.

For sure, the RAW file does not describe noise but the light conditions it
captured. As you see the rendition on your screen, if the shot was under
exposed and one may not like it, boosting the exposition will generate
as much noise as the under exposition will require.

That is because in an under exposed take, some necessary information,
that would make the shot properly exposed, was not yet registered when
the second curtain closed. If you're trying to get an image out of no or
insuffisant information, the software will try to increase details that do
not exist and so… generate noise.

[Colin Southern]

I have been wondering for a while whether noise in an image would be less with a longer exposure time for the same total amount of light.

In a recent thread, if I understand correctly, Colin Southern suggested that noise could be reduced with a large number of shots of the same scene stacked in PP.

I did some experiments. I photographed a blank white wall at f/2.8 SS1/1000, f/5.6 SS1/250 and f/11, SS1/60, all at ISO 1000, with my 100mm macro lens. The total exposure came out at about the same in each case as expected. There was no discernible difference in the amount of noise in the three cases.

This indicates to me that the noise does not get averaged out with time and makes me wonder what causes noise. I would have expected thermal/electronic noise to average out. Is it variations in the pixel characteristics?

Can anyone enlighten me on these questions?

There's many different types of noise, but mostly what you'll hit is due to trying to work with too little signal (ie too close to the sensor noise floor).

If you want to see how it averages out, take 32 shots at ISO 3200 - average them - then look at the noise; it'll be close to what you'd get at ISO 100.

For long exposures (multi-minute) noise increases - shorter multiple exposures averaged reduce this too.

[dubaiphil]

In a recent thread, if I understand correctly, Colin Southern suggested that noise could be reduced with a large number of shots of the same scene stacked in PP.

I'm not going to get all techie with the noise debate, but stacking has one serious advantage.

In low light, when using this technique with star trails for example, it is good practise to capture a dark frame for noise reduction when stacking the files. This dark frame is taken with the lens cap on and the same exposure settings as all the individual shots taken during the shoot. With long exposure noise reduction in your camera switched to off during the sequence.

With long exposure noise reduction set to ON in camera, this is exactly what your camera does by itself, with in camera software comparing exposure to dark exposure

So if you're taking 2 hours worth of images for star trails you have several options, polar opposites being:

1 x 2 hour exposure, with another 2 hours in camera noise reduction
roughly 240 x 30 second images, with another 30 second dark frame

So you're going to be eating up a lot less battery and seeing more instant results with the stacking method.

[Colin Southern]

I'm not going to get all techie with the noise debate, but stacking has one serious advantage.

In low light, when using this technique with star trails for example, it is good practise to capture a dark frame for noise reduction when stacking the files. This dark frame is taken with the lens cap on and the same exposure settings as all the individual shots taken during the shoot. With long exposure noise reduction in your camera switched to off during the sequence.

With long exposure noise reduction set to ON in camera, this is exactly what your camera does by itself, with in camera software comparing exposure to dark exposure

So if you're taking 2 hours worth of images for star trails you have several options, polar opposites being:

1 x 2 hour exposure, with another 2 hours in camera noise reduction
roughly 240 x 30 second images, with another 30 second dark frame

So you're going to be eating up a lot less battery and seeing more instant results with the stacking method.

Another thing to keep in mind is that thermal noise is pretty much just proportional to temperature - so one can prepare a number of dark frames (at various temperatures) in advance.

Probably pays to mention too that star trails are a little different from regular stacking in that you're dealing with a moving subject so the exposure doesn't change; a 240 x 30 -v- 2 hour exposure for a non-moving subject would require vastly different iso/aperture settings.

[DanK]

Noise is generated by the sensor, not by software (although software can exacerbate it), and like any information generated by the sensor, it is in the raw file. It is generated by different processes, but for purposes of your question, these create two basic types of noise: what is usually called fixed pattern noise, and various types of random noise. Here is my understanding of this:

Fixed pattern noise arises because of differences in the behavior of the photosites on a sensor. Except for hot pixels and dead pixels, most of these variations are too small to be seen in a well-exposed image with a short exposure time. In the case of long exposures, however, as in night photography, these become more apparent. Because they result from attributes of specific photosites, they will be in the same location in each image. Therefore, averaging wouldn't help. The solution is subtractive noise reduction: take a dark frame, calculate where in the frame the artifactual light pixels are, and subtract them from other frames. This is what long-exposure noise reduction does, and it is why it locks the camera for a period identical to the length of the original exposure.

In contrast, most noise has a large random element. Random variation is attenuated by averaging. Therefore, in those cases, averaging should help. This is exactly the same principle as the fact that large polls have less variable results than small polls.

Any time you amplify the output from the sensor, you will amplify both signal (in the sense of information) and noise. This is why high-ISO shots have more noise than low-ISO shots (a higher ISO setting tells the camera to amplify the output of the sensor), and why underexposing an image and increasing its brightness in software will make noise more apparent.

Different software tools do this differently, so their effects may vary. In addition, there is an argument about the relative effects of increasing ISO vs. amplifying the signal in software, with most people claiming that the exacerbation of noise is generally worse if you do the amplification in software. It would be easy to test this, but I never have.

I have been wondering for a while whether noise in an image would be less with a longer exposure time for the same total amount of light.

Generally, the reverse is true. If the longer exposure is long enough, you will get visible pattern noise. In addition, particularly in hot weather, you will get more noise as the sensor heats up. A longer exposure is not averaging. one would average across different exposures.

[ajohnw]

The problem with averaging noise is that there are several types. The whole idea is based around noise being completely random so if an infinite number of shots were averaged mathematically speaking it would all disappear. This doesn't apply to exposure time. All that happens is the state of the random noise at that point in the exposure will be captured.

Actually there are other things which will build up as the exposure time is increased. This is what dark frames can be used for. You might regard this as electrical leakage into a pixel that has the same effect as exposure to light. Sometimes when pixel counts are mention on a sensors there are 2 figures, one is larger than the other, The pixels in the difference are never exposed. These can be used for something akin to a dark frame to allow the leakage to be "subtracted" from all pixels that are exposed.

As Colin mentions if a significant number of pictures are averaged for noise removal it will be reduced. How much will depend on how many and probably the camera as well. Higher ISO settings = more noise.

Star exposures can be a bit complicated to get right. An exposure time often has to be set to limit the degree of trails that are produced because they are always moving due to the rotation of the earth. The maximum exposure time decreases as the focal length gets longer. Aperture will often be maximum to get as much light in as possible. That leaves ISO as the only control. The idea here is to set it at a level where the sensor can capture significant levels of star light at the desired level, stars have various brightness. The signal captured in the sensor has to be greater than the noise the ISO level produces. One way of handling this is to take dark frames with the same exposure time and subtract them and then additively stack them if that's needed. Or take many and noise average them or some combination of the 2. Averaging to remove noise is probably more common which means the signals generated from the light must be significant. They generally are so the problem comes about via trying to capture the milky way rather than some group of stars which will be a lot brighter. It needs higher ISO settings as aperture and exposure times are fixed.

Exposure times can also be extended by stacking lots of short ones. The problem of capturing the light levels that are wanted rules here as well. I don't think anyone does this to produce star trail shots.

There is plenty of info on this subject about in astro photography circles on the web. Another type of control exposure that is often used there is flats.

John
-

[Colin Southern]

Noise is generated by the sensor, not by software

Sorta / Kinda.

Software gets involved though; consider a high ISO shot that's been under-exposed. Because it's under-exposed the signal to noise ratio is poorer than it needed to be, but if you chuck the image straight out of the camera without any adjustments you probably won't see the noise ... but you won't see an ideal photo either - so - we adjust the image using software to compensate for the under-exposure and THAT'S when we see the noise.

So from that perspective the software processing doesn't "cause the noise" per se, but it does "reveal" the noise.

[dje]

Tony you might find this article useful. It gives a fairly in-depth discussion of noise in digital cameras.

Dave

[DanK]

So from that perspective the software processing doesn't "cause the noise" per se, but it does "reveal" the noise.

Yup. That's a more concise and clearer alternative to what I wrote:

Any time you amplify the output from the sensor, you will amplify both signal (in the sense of information) and noise. This is why...underexposing an image and increasing its brightness in software will make noise more apparent.

However, it's not necessarily the case that you need amplification to make the noise apparent. My favorite example was one of my first tries at wilderness night photography. I tried a very long exposure--I think it was an hour or two--on a very hot, humid night, using a Canon 50D. You can guess what happened. When I opened the image in LR, I was horrified--a huge amount of noise, weird but very pronounced irregular color banding, etc. It was utter garbage. I'm assuming a good bit of it was thermal noise, but whatever the process, abused in that way, the sensor produced so much distortion that no amplification was required at all to make it glaringly apparent.

[Colin Southern]

However, it's not necessarily the case that you need amplification to make the noise apparent.

Yep.

To a certain point you can often control the noise with an aggressive black clipping point, but past a certain point (when we start talking hours), digital suffers - especially at high temps.

I'm sure the astro boys have better solutions (nitrogen cooled sensors etc), but that's waaaay outside my areas of expertise.

[TonyW]

Tanks for all the replies to my initial questions.

Dave, your link was especially useful. It seems that there are two classes of noise. There is the "pixel response non-uniformity" which causes noise consistently over different pixels. I would not expect this to average over time or over several shots and is perhaps what I was observing. This type of noise is more likely at relatively high exposures.

The other class is the random noise over time which includes sensor read noise, thermal noise and quantisation errors due to the conversion of more or less continuous values of voltage to discrete digital values. These types of noise i would expect to average over time or multiple exposures. I still don't see why noise reduction by averaging over multiple exposures would work where averaging over a longer exposure would not, unless it is simply because a longer time exposure heats up the sensor and produces morel thermal noise. Noise is a problem normally only when the exposure value is low so that the thermal effect would be small.

In any case, as suggested already, the best way to deal with noise is to avoid it in the first place if that is possible.

[DanK]

I still don't see why noise reduction by averaging over multiple exposures would work where averaging over a longer exposure would not,

Perhaps someone with more understanding than I have of the electronics can explain more, but I think the main issue is that longer exposures do not average; they simply accumulate. Once a pixel is illuminated, it doesn't fade. In contrast, when you average exposures, you are taking the mean illumination of that pixel.

[HaseebM]

Yep.

I'm sure the astro boys have better solutions (nitrogen cooled sensors etc), but that's waaaay outside my areas of expertise.

Yes, astrophotographers do utilize cooling technology to their DSLR's while some prefer cooled dedicated imagers. However this is not really a worry for astrophotography as for the amount of light frames they shoot, they also take Dark frames at same ISO, Shutter as well as Offset frames which are taken by using the fastest shutter but same ISO. All these frames are then combined in stacking software and then stretched using levels / curves and other tools like Noel Carboni's astro tools for eg. For those who get too much vignetting, they take 'flat frames' which are shot using 1/200, aimed at twilight sky or lightbox.

[ajohnw]

Some do more than that. I'm not up to date on it but some were looking at removing the bayer layer from the sensor. Cooling dslr's has also been done.

The basic aim is to be able to capture the lowest level of light possible. The reason is pretty simple. When a telescope is used it has a greater orifice ( can't use aperture on here) than our eyes so captures and concentrates more light into our eyes.so dimmer stars can be seen. Sensors are more sensitive to light than human eyes so can see even more through any size of telescope. This has had an interesting effect. Something like a 75mm telescope is visually a bit limiting - it way way better with some sort of camera on the end but anything that that makes them more sensitive or reduces noise makes this aspect even better. The pro's do use gas cooled sensors - I believe they manage to get close down to absolute zero. Cooling does have it's problem. Condensation so something like a vacuum flask has to be put in the way. The other odd aspect is that metal objects cool rapidly under clear night skies. That causes condensation problems as well.

John
-