All you need to know about noise
It's well known that images from digital cameras show higher noise when using higher ISO settings, and that small digicams also show higher noise than DSLRs. As a photographer you can accept this as a fact, but if you're technically interested the following questions may come up:
- Why is there more noise at higher ISO settings?
- Why is there more noise with smaller digicams?
- Why is there more noise with longer exposures?
- What can be done against noise?
Short introduction to image sensors
In the old days, our image sensors were composed of light-sensitive particles, embedded in gelatine, pasted onto a cellulose triacetate carrier. We called them “film”. Light falling onto these particles caused a chemical reaction, and it was amplified and fixed during processing.
Digital image sensors work similar. There you have light-sensitive, isolated areas on a silicon carrier, called photo sites. When light falls onto these during exposure, electric charge builds up in the photo sites through the photoelectric effect. After exposure, the charge of each photo site is read out, amplified, digitized and stored on a storage medium.
The problem with these photo sites is that charge is not only caused by light falling onto them. Other factors also cause charge buildup, mostly temperature, but also cosmic rays, radioactivity, electricity leaking from neighboring photo sites, etc., and the amplifiers and A/D converters introduce their own noise. This minor part is – on average – more or less constant, i.e. not depending on exposure, and – looking at a single photo site – random. For example, a photo site may be hit by a radioactive particle during exposure, or it's not. This part of the total charge is what we see as noise.
Higher noise at higher ISO
So if noise is constant, why do we see more noise at higher ISO settings?
Now here's a fact that may surprise you if you did not yet dive into this topic before: You can not change the sensitivity of an image sensor. The sensitivity is fixed when the sensor is designed and built, determined by the layout of the photo sites, the materials used and by support structures like microlenses. Changing sensitivity means building a different sensor.
So, you might ask, what is the ISO setting of your camera doing if it's not changing sensitivity? The ISO setting is doing two things:
- It changes exposure metering, meaning that at higher ISO settings the camera chooses to expose the sensor less, and
- it cranks up the signal amplifiers that sit between the sensor and the A/D converters, to compensate for the smaller exposure.
Of course, the amplifiers can not distinguish between the charge of the photo sites caused by light and that caused by other sources, so it amplifies both. Since noise is more or less constant in absolute terms, it grows relative to the desired signal.
This is also illustrated in the following picture:
On the left side we see the charge of a photo site. The red part is noise, caused by sources mentioned earlier, and the green part is the signal caused by photons, i.e. light. The middle column is the same situation, only at a higher ISO setting. Noise is the same, as mentioned before. However, the signal is smaller, because at higher ISO settings exposure is reduced by the camera. To get the same apparent brightness, the amplifiers compensate for reduced exposure, increasing both the desired signal and noise.
In real life, noise is of course a lot lower compared to the signal. The illustration is only qualitatively correct, not quantitatively.
Higher noise with smaller sensors
With smaller image sensors, like those used in small digicams and cell phones, you also get higher noise. This is not because noise is inherently higher with these. These small sensors contain also relatively small photo sites. The smaller a photo site, the lower its sensitivity. So to reach their nominal ISO rating these cameras have to amplify the signal from the sensor more than other cameras with larger photo sites. This higher amplification is what is causing higher noise.
There are also minor factors responsible for higher noise. Typically the signal processing in these cameras is less sophisticated, and additional noise is actually coming from the amplifiers and A/D converters instead of the sensor.
Higher noise with longer exposures
If you compare photos taken with long exposure times to photos taken with short exposure times, you also often see higher noise. This is true even though total exposure of both is identical and the same ISO setting was used. Why is this?
The reason is again the different sources of signal and noise. The signal is caused by light falling onto the photo sites, and the charge caused by this is the same, no matter if you expose briefly with much light or for a long time with little light. But noise is not caused by exposure but by independent sources. The longer the exposure time, the more opportunities for these other sources to cause noise. A 2 second exposure contains twice as much noise as a 1 second exposure, because temperature etc. can influence the photo sites twice as long.
How to avoid noise
If you want to avoid noise the first thing to do is use a camera equipped with a sensor having a high native sensitivity (typically, a DSLR with a large sensor and large photo sites), and use a low ISO setting. This keeps amplification at a minimum.
You can also use noise reduction, if your camera supports it. With this feature the camera first makes the intended exposure, followed by another exposure of the same length, but with the shutter closed. The second image contains only the noise that occurs at the current temperature with the current sensor. The data of the second image is then subtracted from the first image, resulting in an image with strongly reduced noise (noise is not zero; because of the random nature of noise, the second image is not 100% identical with the noise of the first image). This noise reduction technique is called “dark frame subtraction”, and many cameras offer it for longer exposures.
You can also use software to suppress noise that still ends up in the image. But suppressing noise instead of avoiding it in the first place, and at the same time not losing image detail is a tricky task. You need sophisticated software and know how to use it well.
Other than this, there is not much the photographer can do.
For sensor makers, there are a few things more that can be done. Thermal noise can be avoided by cooling down the sensor. However, this is not practicable for photo cameras, but only for sensors attached to telescopes etc. But camera makers can at least avoid heating up the sensor too much.
Sensor makers can also increase native sensitivity, on one hand by making photo sites large, and on the other hand by increasing quantum efficiency, i.e. increasing the fraction of photons actually causing a charge increase. But quantum efficiency is already high, around 70%. So even if quantum efficiency is increased to 100% (which is unlikely), sensitivity can at most increase by 1/3. And photo sites tend to shrink instead of grow, because of the never-ending demand for higher pixel numbers.
Sensor makers can also improve their signal processing chain, making sure that noise introduced there is as small as possible. Moving towards CMOS sensors instead of CCD sensors is one way, because with CMOS sensors the processing electronics is closer to the individual photo sites.
References and links
For more detailed discussion of selected topics, please refer to these other web pages:
© 2007 Michael Hohner; This page was last changed on 2007-06-02
|#1: Comment posted by Donca on August 22nd, 2008 - 09:24:31 AM:|
|I don't have a great technicall knowledge, but after my experience taking photos in concerts, I have realized that noise is produced, above all, in the subexposed zones of the photography.
And with higher ISO but higher EV, the noise is lower. And it is easier to correct it in the post-process of the photography.
|Michael Hohner answers:|
|In absolute terms, noise in areas of low exposure is the same as noise in areas of high exposure. It's only that you can see noise more easily in the areas of low exposure because it's large relative to the signal.|
|#2: Comment posted by scott on November 10th, 2008 - 07:06:44 AM:|
|What about for longer exposures? My reasoning would suggest an asymptotic noise/signal relationship for very long exposures.
I know if I use a 3200 ISO for 1/32 second and a 100 ISO for 1 second then I'll get a much less noisy image (for the 100 ISO).
Can you elaborate on how noise gets into the system for longer exposure times?
|Michael Hohner answers:|
|I have written about long exposure noise above. Have you read it? Basically, noise is higher for longer exposures at the same ISO setting. If you compare different exposure times at different ISO settings, noise decreases with lower ISO settings and increases with longer exposure times. There is a point of lowest noise, but where that point is depends on the implementation.|
|#3: Comment posted by rob on July 29th, 2009 - 08:43:45 PM:|
|I have tried the following to try to reduce noise when shooting with a Sony DSLR at high ISO (800) but the results are incnclusive as yet - I have to do more work on this one.
First, when shooting at ISO 100 I shoot in RAW, but when I have to up the ISO I switch to JPEG and 5Mp size instead of 10Mp - my reckoning is that if some noise is due to "overpopulation" of the CCD, then effectively reducing the pixel count might lead to less noise.
I haven't done any detailed trials on this but I have to say that the noise appears to be better - this could be simply because I want it to be better. As I say, it needs thorough trials before I can say it works, but it's worth a go.
Anything is better than noise reduction in photoshop!!! Even noise reduction in Adobe Camera Raw is pretty poor.
|Michael Hohner answers:|
|Shooting at a lower resolution will reduce noise if, and only if, the camera reduces the pixel number by interpolating neighboring pixels. The signal of neighboring pixels is likely very similar, and noise is likely different. So the random parts, ie. noise, will likely cancel out each other, and the signal parts will be enforced.
If, however, the pixel number is reduced by just leaving out interleaving pixels, noise will not be reduced.
|#4: Comment posted by Andres on September 27th, 2009 - 09:41:58 PM:|
|When photographing still subjects in low light I always wondered if boosting ISO would be better than a longer exposure or is it the other way around. For instance ISO 200 2 second exposure versus ISO 400 1 second exposure. One would get less noise amplification from ISO and more from the other "time dependant" sources you mention for longer exposures. I tend to think that the camera can handle noise from long exposure better than from amplification (boosting iso) because it does this closed shutter background noise substracting. What's your take on this?|
|Michael Hohner answers:|
|I think this can not be resolved by theory. Too much depends on the actual implementation. It has to be tested with the specific camera.|
|#5: Comment posted by Bogdan on March 24th, 2010 - 11:01:49 AM:|
THANK YOU for a very interesting and instructive article.
Not only that it looks like you know what you speak about, but you have the ability to make other to understand what is all about. And this is a very rare quality. I have met very many experts in one or another area who haven't a clue what communication is about and therefore are unable to make their messages through. This is definitely not your case.
Thank you very much in deed for your article. I'll come back and check out the rest of the site.
|#6: Comment posted by Pranav on January 31st, 2012 - 10:39:01 AM:|
|Is it true that size of photosite and not the sensor size is responsible for avoiding noise?|
|Michael Hohner answers:|
|Yes, as explained above.|
|#7: Comment posted by Urban Domeij on March 28th, 2012 - 02:04:53 PM:|
|I think the answer to #6 is incomplete.
In any image that is viewed in real life, individual pixels are mostly interpolated, thus in many cases canceling out noise, while reinforcing signal. The same applies when binning pixels in a camera, as the large format cameras when bundling four pixels, effectively causing a 65 megapixel back to register 16 megapixels. Noise is suppressed to some degree, causing the bundled photosites to act as were they larger but fewer photosites.
Therefore, more pixels in the same area in fact will not increase noise, although more noise may be present in some pixels at pixel level. A similar effect will be experienced when interpolating an image to a smaller size, like in some of our compact cameras. Apparent noise is decreased, due to dissimilar levels of noise cancelling each other, but similar levels of signal enhancing signal to noise ratio. Thus an interpolated but shrunk image will be less noisy than the original raw image from the sensor when viewed at pixel level. The same thing may apply when making a real life picture on screen or paper by interpolation to adapt it to size and printing method.
Of course we may reach a point where it does not make sense to make photosites smaller, because they won't enhance lens resolution, but I don't see any disadvantage regarding noise or sensitivity. Bundled photosites are more sensitive to light than single photosites, because the four photosites receive four times as much light with the same exposure, while noise is only about twice, due to some noise factors cancelling out each other. So the determining factor for sensitivity and noise is in fact not photosite size, but sensor area. Photosite size may determine resolution though, which makes it more important to shrink photosites in small sensors than in large sensors, and will always cause small sensors to produce images with less dynamic range than larger sensors, provided they use the same technique. Increasing dynamic range in a small sensor will cause loss of resolution, so some balance will be found by designers, in order to have sufficient of both, leaving some of the decision to the photographer, who may sacrifice excellently low noise at low ISO settings against higher noise and lower dynamic range in low light by amplifying signal and noise together. At some level, bundling photosites by interpolation will make sense.
The bottom line is that it is not primarily photosite size, but sensor size that matters for dynamic range and actual image resolution.
|Michael Hohner answers:|
|Pixel binning is not considered here. But remember that even with pixel binning, noise will likely be slightly higher (depending on the shape of photosites). For example, if you replace 1 large site with 4 small ones, you will likely not get the same total sensitive surface area because of the required isolating borders between the pixels.|
|#8: Comment posted by fan on December 11th, 2012 - 06:09:10 PM:|
|Great, learned a lot.