Understanding how your sensor responds can open up creative options that aren't otherwise obvious. For instance, this high dynamic range scene was achieved using a single exposure, thanks to the camera's excellent noise characteristics.

In the first part of this article, we discussed the fact that a lot of the noise in a lot of your images doesn't come from your camera at all: it is you being able to see the randomness of the light that you've captured and it is almost solely dependent on how much light you were able to capture. The second source of noise is more likely to resemble what most people think of, when they think of noise: it's the background electronic noise that's added to your images by the specific camera you're using. We say this is the source of noise that most people think of because it's much easier to think of noise the way you think of background 'hum' from a Hi-Fi amplifier.

The key thing to remember though is that, although shot noise and electronic noise stem from very different sources, with the exception of a few special cases (such as 'banding' pattern noise), they are visually indistinguishable. Noise is just variation from the expected result: aberrantly brighter or darker pixels or spurious color.

Just like with a Hi-Fi amplifier, you'll recognize that noise tends to be easier to hear during quiet patches in the music when you've got the volume turned up. In a similar manner, the electronic 'read noise' rumbles along in the dark tones of your image (where there isn't much signal) and is most likely to be visible when you've pushed your signal a lot (you're shooting at high ISO).

To better understand the behavior of different sensors, we're going to split the sources of electronic noise into two groups and consider them separately. We're going to draw the line between noise added before amplification and any added afterwards. These come in addition to light's inherent shot noise that we covered in part one.

What we're going to call Upstream read noise includes the electronic noise contributed by the pixel during the process of image capture*1. Downstream read noise includes the contribution of all the downstream electronics and all other sources after ISO amplification*2.

The reason we'll look at these two sources separately is that, while they tend not to change very much (for any particular sensor), the upstream read noise has already been added to the signal before amplification, so gets amplified along with the signal when you change ISO, whereas downstream read noise stays the same, regardless of amplification. This difference becomes important when you compare performance between sensors, as it can dictate not only the results they give but also the most effective ways to use them.

Most modern sensors have pretty low upstream read noise, the big differences between different sensor designs tend to be in how low the designers have managed to get the downstream read noise. For the rest of this article we're going to compare the behavior of two theoretical cameras: one with low upstream and low downstream read noise, and a second that differs by having higher downstream read noise, to see how and where their performance differs and consequently, how you should shoot with them. We've going to use the same tone flow diagrams that we used in part one of this article, tracking what happens to the tones in the scene, from the point of capture all the way into the final image.

At low ISOs

At base ISO, where there's no ISO amplification being applied, the low upstream read noise makes very little difference. However, the differences in downstream read noise can be seen in terms of low ISO dynamic range: a camera with lower downstream read noise will have greater low ISO dynamic range.

There's every chance you won't see this in the camera's JPEGs, because even the noisiest contemporary cameras tend to have a noise floor below the darkest tone used in the JPEG. However, the best modern cameras can have several stops of additional usable dynamic range that you'll simply never see in your JPEGs. This is because JPEGs tend to have tone curves designed to create punchy images with plenty of contrast when displayed on monitors or as prints (which have relatively narrow dynamic ranges), but it means there are a lot of creative possibilities that open up if you want to shoot Raw and make use of this extra information (such as the image at the top of this page, which looked like this prior to a custom tone curve to bring out shadows).

In this diagram, pushing the Raw file to pull more deep shadow information into the image (by applying a different tone curve), starts to reveal the difference in noise performance between these two sensors. On a camera with low downstream read noise, you still won't encounter its effects, even after a substantial pull from the shadows. Hover over the right-hand tab and you'll see what happens if you have higher downstream read noise: that noise gets pulled into the image.

This situation is made even worse when the downstream read noise includes a source of noise that introduces a pattern to the noise. Banding is a particular problem: because it's not totally random, it doesn't get cancelled-out by re-sizing images. Worse still, humans are especially good at recognising edges, so will tend to see patterns and find the noise much more distracting than totally random noise from other sources.

This flexibility is also valuable if you or your camera ever accidentally underexposes an image. For example, with a back-lit subject even the best metering system can sometimes get distracted by the bright background and lower the exposure. This will increase shot noise but a camera with lower downstream read noise will provide a Raw file that is more tolerant of being brightened. Using such a camera should contribute less noise than is evident in the groom's jacket in the example below.

Corrected image, showing additional noise and a limited tolerance for exposure correction. Original image, Adobe Camera Raw default processing.

At high ISOs

Having said that most cameras have low upstream read noise, it starts to play a bigger role at higher ISOs, where it gets amplified, along with the signal (and all the shot noise it already contains). If we assume that a camera has a base ISO of 100, then ISO 12,800 will see any upstream read noise (and shot noise) amplified to 128 times its original, base ISO level. This will significantly dwarf any noise contribution coming from downstream read noise; so a camera whose high downstream read noise limits its low-ISO dynamic range could still have excellent high ISO performance (where shot noise and upstream read noise tend to be the limiting factors).

Because this is a high ISO example, it uses a shorter exposure and more amplification. This can be seen in the diagram, where the scene brightness is captured as a series of darker tones on the sensor, then pushed back up by amplification before being written into the Raw file. The shorter exposure means less light is captured and all the tones from the scene are represented by small signals with poor signal to noise ratios (because of shot noise). The darker parts of the image have very little signal already which means that not only is shot noise already a large proportion of the signal, but also that even small contributions of upstream read noise will have a big impact on the signal-to-noise ratio.

The amplification means that these poorly-recorded signals (and the read noise added to them) end up as relatively bright, visible tones in the final image. However, because the downstream read noise is not amplified, its role is insignificant in comparison.

Note that the brightest tone in the Raw file and in the final image wasn't the brightest tone originally captured by the sensor. Brighter tones were captured with this exposure but then were amplified to the point that they can't be fitted in the Raw file: they've clipped and been recorded as overexposed white. Also note that all the data used to make up the final image is taken from values above the downstream read noise floor in both cases.

This means two things. Firstly, that on cameras with low downstream read noise it can make more sense to leave the ISO setting low, because you won't add significantly more noise by boosting the brightness with a tone curve, rather than by using ISO amplification, but you will retain more highlight data. By contrast, on a camera with higher downstream read noise, increasing the ISO amplification can push all your image data so that it ends up above the level of this camera's downstream read noise. This give a result that's less noisy than shooting at low ISO and brightening the image later, because brightening the dark tones would effectively amplify the downstream read noise.

Clarifying the meaning of our Raw Dynamic Range tests

These articles should also help you understand what our Raw Dynamic Range tests are showing. Our Exposure Latitude test illustrates what happens if you try to push a low ISO Raw file: a very obvious simulation of what you might do in the real world. However, it has the disadvantage that, because we have to reduce the exposure between images, we introduce more shot noise, so it's not possible to tell how much of the additional noise we see is coming from shot noise and how much is coming from read noise. As a result, you can only assess a good sensor performance by comparing cameras of the same sensor size (which will experience the same amount of light*3, meaning the differences are predominantly the result of read noise).

To solve this problem, our ISO Invariance test uses the same exposure at each setting, which means any noise differences must come from (electronic) read noise. This allows us to take a closer look at the camera's read noise contribution, without it being complicated by the effects of shot noise. This should make clear that, although it may initially seem a little unconnected to the way you shoot, it actually gives a useful insight into a camera's performance.

From this you can work out the extent to which you can choose a lower ISO setting (though still use a short exposure) and push tones selectively later, getting the same amount of noise while retaining more highlight information.

Image shot at ISO 3200, 1/100th, F8 to achieve appropriate subject brightness. Adobe Camera Raw default conversion.  Image shot at ISO 100, 1/100th, F8 pushed to match subject brightness but retain highlights. Even using the extensive highlight recovery, the sunset can't be recovered in the ISO 3200 shot.

Take home messages (TL;DR)

The main reason we wrote these articles is to address common misunderstandings about noise. As ever, the full story is a little more complicated than this. You'll hear more about quantization error in a forthcoming in-depth piece about ISO Invariance, but if you'd like to learn more about the specific sources contributing to read noise*4, then Emil Martinec's summary is a great place to continue your reading. However, we're hoping these two articles have made a couple of things clearer:

Firstly: the source of noise you encounter will depend on where you look in your image and how you shot it. Most of the noise you encounter wasn't contributed by your camera: it was shot noise from the light you captured and is primarily dictated by shutter speed, f-number and sensor size. Not pixel count.

Where your camera is contributing noise is in the darkest captured tones. At low ISOs, a camera with low downstream read noise will give you more flexible Raw files, allowing you to capture much more dynamic range than your JPEGs, opening up creative possibilities and giving files that are more tolerant of accidental underexposure.

At higher ISO, the upstream read noise and shot noise are increasingly amplified, so will appear in ever brighter tones in your image. This amplification, combined with the tiny amounts of light being captured means any small differences in upstream read noise or sensor efficiency differences will become exaggerated.

To find out how your camera behaves, our ISO Invariance tests try to show when you can expect to encounter read noise. This not only tells you how flexible your Raw files will be, but also lets you know whether a camera gives better results by upping the ISO or by keeping the same exposure values, lowering the ISO and boosting the brightness selectively, later.

Ultimately, though, we're hoping that with this information you can understand why some cameras produce less noisy images: larger sensors tend to allow more light capture, which minimizes shot noise. However, the differences in sensor performance can be enough to wipe-out these sensor size differences - especially in terms of low ISO dynamic range.

Part 1: the role of shot noise

With special thanks to Prof. Emil Martinec, Prof. Eric Fossum and Iliah Borg


Footnotes

*1 The most salient noise sources upstream of ISO amplification include reset noise, pixel amplifier noise, and dark current shot noise for longer exposures (which we won't dive into). Reset noise originates from variations in the voltages each pixel is reset to after a charge is read (ideally, they'd all be reset to the exact, same voltage). These variations - that show up as changes in intensity across neighboring pixels - can largely be mitigated by a process called correlated double sampling (CDS), which samples and subtracts the reset voltage at any pixel from the total voltage due to exposure. It can almost be thought of as a dark frame subtraction, and is very effective, lowering pixel-to-pixel variations, as well as pixel-level amplifier noise, to the level of single electrons in modern sensors. However, CDS isn't completely effective, and so a non-zero noise component still remains. For reasonable exposure lengths, any remaining reset and pixel amplifier noise post-CDS are probably the largest sources of upstream read noise. [Jump back to text]

*2 Downstream read noise includes all sources from and after, or 'downstream of,' ISO amplification. This includes noise from the programmable gain amplifier (set by your ISO setting), noise from the analog to digital conversion process, as well as any noise introduced in the pathways between all these electronic components in your camera. [Jump back to text]

*3 How much of this light a camera is able to make use of depends on its sensor efficiency, which is beyond the scope of this article. We mention it here, though, because sensors with higher efficiencies will record more of the available light during any given exposure: leading to less shot noise and a slightly cleaner result in our exposure latitude test. [Jump back to text]

*4 There are other sources of noise that are beyond the scope of this article. Thermal noise builds up with long exposures, amplifier glow can lead to 'hot spots' at the edges of your image, pixel-response non-uniformity can lead to noise even in brighter areas of your image, and quantization error further limits signal-to-noise ratio in shadows. [Jump back to text]