20
\$\begingroup\$

Is there a possibility that future tech advances may reduce or eliminate the noise when using high ISO setting, or is this noise unavoidable and inherent to all digital sensors?

If there is a theoretical limit where noise is inevitable, how close are we to that?

\$\endgroup\$

5 Answers 5

27
\$\begingroup\$

It's very important to realize that it is not the high ISO setting itself that results in noisy image, it's that fact that using a high ISO setting means you capture very little light.

Light is made up of photons which are randomly emitted by a lightsource. When the light levels are low or the exposure time very short then the number of photons you get will vary considerably from

Imagine you are trying to estimate the rate at which people are leaving a shopping mall. If you only have 10 seconds to count people then the result you get will vary a lot depending on exactly when you start counting, and which exit you chose. If you have 10 minutes to count people, then you will get a much more stable answer which will be similar for all exits (assuming there is no personal preference for exits) and across different 10 minute time windows (assuming there are no other factors influencing the result).

That is what is happening when you use a high ISO setting, you capture very few photons so a set of neighboring pixels covering an object of uniform colour might receive 4, 3, 4, and 5 photons each, so instead of a smooth uniform colour you get a grainy result that changes for each pixel.

This noise is called photon noise and is the dominant source of noise in high ISO images except in the shadows. Even if you had a perfect sensor that counted and faithfully reported each photon that hit the sensor you would still have a significant amount of noise in low light.


That's not to say that we have reached the limit of high ISO performance. Not quite yet any way. Pure photon noise is very fine grained is less objectionable than the clumpy pattern noise observed in high ISO photographs.

Reducing pixel cross talk, improving the electronics in general might only have a small effect in reducing noise amplitude, but a larger effect on improving noise quality.

Wikipedia has a simulation of the "perfect" sensor where photon noise is only noise source:

Click for a larger version where you can make out individual pixels. Image by Mdf some rights reserved.

\$\endgroup\$
5
  • 2
    \$\begingroup\$ This is true for very short exposure times, but how short? Can you add some (estimated) values for the different exposures in the example picture. Are we talking about 1 nano second to 10 nano seconds, or are we approaching "normal" camera performance exposure times? I realise this will depend on the amount of light, but take a "normal" indoor scene as an example. \$\endgroup\$ Mar 17, 2014 at 16:25
  • \$\begingroup\$ I like this answer, but ` you capture very few photons so a set of neighboring pixels covering an object of uniform colour might receive 4, 3, 4, and 5 photons each` - aren't we still talking millions of photons? \$\endgroup\$ Mar 17, 2014 at 21:48
  • \$\begingroup\$ @KirkBroadhurst That's the whole point: in low-light we aren't. Human vision is approximately logarithmic, and the "stop scale" is also logarithmic. One stop less light means half as many photons. If you start halving, you get to only a few photons very very quickly. If you're mathematically oriented you might want to read up on the Poisson process. Generally, if you have k photons on average per pixel, the magnitude of the pixel noise will be sqrt(k). \$\endgroup\$
    – Szabolcs
    Mar 17, 2014 at 22:59
  • \$\begingroup\$ @KirkBroadhurst Historically, the first model of light was as "rays" (geometrical optics). Then came wave optics. Then quantum mechanics---light is made of discrete units. It is interesting to think that phenomena related to each model (and not explainable by previous ones) do have practical significance in digital photography. \$\endgroup\$
    – Szabolcs
    Mar 17, 2014 at 23:02
  • \$\begingroup\$ @Matt Grum - The second paragraph appears to be incomplete. \$\endgroup\$
    – Michael C
    Jul 8, 2014 at 23:29
11
\$\begingroup\$

Reduce it, yes. For example, the Canon 5D Mark III is 2/3 stops better than the Canon 5D in high ISO performance, although their sensors are the same size, because it is seven years newer. Of course, past performance is not necessarily indicative of future results, but I see no reason for incremental gains not to continue to be made.

Eliminating it completely is physically impossible. When you get to an ISO in the millions, you're trying to extract data out of a few photons. Regardless of how good your technology is, the information is simply not there for you to extract.

Now, as for getting it "perfect" for all ISOs under, say, 3200, note that there isn't really a consistent standard for "perfect." You might develop some amazing new technology that reaches some theoretical bound in signal-to-noise ratio, but does that really matter when my eyes claim this pixel should be #0f3ed2, you claim it should be #0e3fd4, and the sensor thinks it's #0d3dd3?

\$\endgroup\$
5
  • 5
    \$\begingroup\$ "Perfect" would be a photon counting sensor with infinite capacity. You could actually do that today (except for the infinite capacity part), but it would be very expensive. But even that would be noisy in low light. The information simply isn't there as you suggest. \$\endgroup\$
    – Matt Grum
    Mar 17, 2014 at 11:06
  • \$\begingroup\$ @MattGrum: What if we could make the sensors sensitive only for a very narrow spectrums, so that it only counts photons of a specific energy? wouldn't that remove most of the noise that in contemporary sensors somes from things like thermal effects? \$\endgroup\$
    – PlasmaHH
    Mar 17, 2014 at 14:16
  • 2
    \$\begingroup\$ @PlasmaHH - you're still stuck with the non-deterministic nature of light. There's no way around that, except to keep your "poll" going for long enough that the statistical sample variation is negligible. Or, in photographic terms, you need a higher exposure to reduce noise. At some point, no matter how efficient your sensor is, you'll be calling too few people to accurately predict the election results, so to speak. \$\endgroup\$
    – user2719
    Mar 17, 2014 at 14:21
  • \$\begingroup\$ @StanRogers: Ah, so you mean the noise that is looking similar to photon tracing images with small sample sets. I was always thinking of noise as additional signal "on top" of the perfect photon counting. \$\endgroup\$
    – PlasmaHH
    Mar 17, 2014 at 14:54
  • \$\begingroup\$ @PlasmaHH Yes, exactly. There simply isn't enough photons (in this case, we can pretend photons really are just some distinct balls bouncing around) to paint an accurate picture. This gets much better if you don't care about color (even more so for human vision which is much better at seeing light intensity), but it still is finite. There's also some noise in the sensor (for example due to photon cross-talks, which is where photons-as-balls breaks), but that's where the limitation is only practical - bigger sensors and better lens eliminate this almost entirely. \$\endgroup\$
    – Luaan
    Mar 18, 2014 at 8:22
8
\$\begingroup\$

It already happened! On film, or early digital, high ISO meant 400, on latest full frame cameras it means 6400. Problem is that each time it happens, 'High ISO' gets redefined to be even higher, or in another words, high ISO always means "so high that current tech makes it noisy". As noted by Tony, there are eventually, physical limitations as to how far it can go.

\$\endgroup\$
4
\$\begingroup\$

Via Hacker News, I recently came across this paper from 2008, written by physics professor Emil Martinec in his spare time apparently.

Noise, Dynamic Range and Bit Depth in Digital SLRs

He characterizes the different types of noise that are possible, and describes their relative importance.

  • Photon shot noise
  • Read noise
  • Pattern noise
  • Thermal noise
  • Pixel response non-uniformity (PRNU)
  • Quantization error.

After reading this you'll realize that it is impossible to entirely remove the various types of sensor noise. Certainly it is possible to minimize them (in various ways), but there also other design decisions that the camera/sensor manufacturer must make that may introduce other problems or trade-offs (e.g. applying offsets in the A/D converter, see Fig. 10+11)

Regarding your questions about a theoretical limit:

"The most important noise sources for typical exposures are read noise and photon shot noise."

"The inverse of the slope of the PRNU graph (see Figure 7 for an example) is an upper limit for the S/N ratio, unless PRNU is compensated for in post-processing."

\$\endgroup\$
-1
\$\begingroup\$

This is a problem with sensors in general - from optical sensors to accelerometers and gyros. All consumer products deal with this and try to hide the noise from the user - for example, your phone is capable of sensing vibrations way below the level that causes it to take action, and there are apps which can show you that.

Any sensor capable of recording signals accurately within the area of interest will also be capable of recording signals outside the area of interest, and signals below or above the threshold of interest are generally called noise. This 'issue' is not related to optical sensors only, it's related to the physical limitations of sensing the things we are interested in.

So the answer is no - any sensor which is 'insensitive' enough to eliminate noise will also eliminate some of the signal we want, making it impossible to build non-noisy sensors.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.