FAQ FAQ  Forum Search   Events   Register Register  Login Login

Topic ClosedIs sensor noise really that important?

Page  123>
Author
Miranda F View Drop Down
Senior Member
Senior Member

Joined: 11 January 2014
Country: United Kingdom
Location: Bristol
Status: Offline
Posts: 2569
Direct Link To This Post Topic: Is sensor noise really that important?
    Posted: 13 January 2018 at 21:27
I notice a lot has been written about sensor noise, and whole websites (like Photonstophotons) devoted to measuring and comparing sensors; many comments get added to camera review websites and discussion forums by people saying they wouldn't choose such-and-such a model because either the noise was too high, or the dynamic range was too poor, or how high-ISO noise was so much better on one camera than another.
But I rather think this is a waste of time, and this is why.

It seems to me that in most real photographic images (that is, not special test shots), particularly high-ISO shots, the majority of the noise you can see in the image is nothing whatever to do with sensor noise and has very little to do with the camera at all, including the sensor, the amplifiers, and the ADC converters. It may have a little to do with the filtering algorithms, if these are used, and it will certainly be affected by some of the camera settings (DR, for example). But if you can turn the filters off and examine either the raw data or the jpegs (straight in-camera or in PP), you will find that the noise is definitely not attributable to the sensor.

To prove this, I have conducted a little experiment in which you can take part. The second post in this thread has some theoretical justification for my argument, the third has a practical test you can do yourself, and the fourth has my test images you can examine. You can take your own images and post them below for comparison, on different cameras/etc.


Edited by Miranda F - 13 January 2018 at 21:54
A900, A58, 5d, Dynax 4, 5, 60, 500si/600si/700si/800si, various Sony & Minolta lenses, several Tamrons, lots of MF primes and *far* too many old film cameras . . .
 



Back to Top
Miranda F View Drop Down
Senior Member
Senior Member

Joined: 11 January 2014
Country: United Kingdom
Location: Bristol
Status: Offline
Posts: 2569
Direct Link To This Post Posted: 13 January 2018 at 21:29
The theory.
The two things which seem to worry us most about sensors are dynamic range and sensor noise. You can read much more about dynamic range on Bill Claffs website (http://www.photonstophotos.net/); indeed probably more than you ever wanted to know.

The highly simplified version is that DR is the amplitude range (in stops or EV, powers of two) between the smallest and largest signals for which the sensor will respond properly This is limited at the bottom end by Quantum efficiency, QE, described below, and sensor noise, and at the top end by full-well capacity (where the sensor saturates). In principle the pixels for different colours could saturate at different levels but experience suggests they are arranged to saturate at the same time so that saturated pixels look white. Experience also suggests that the red channel shows more noise than the green (which you would expect since there are more green pixels) and also more than the blue.
The more recent Sony sensors have a higher QE (so they are more sensitive) and possibly also a higher full-well capacity (ie they saturate at a higher light level), giving increased DR over older models like the A850/900.

There are three main sources of noise in digital camera images. These are:
1) Thermal noise (kTB). All electronic circuits suffer electrical noise caused by thermal vibrations in the conductor. You can read more about it in wikipedia, but essentially the noise is proportional to absolute temperature (kelvin), and to signal bandwidth. In a well-designed camera, thermal noise can be kept at a low level and shouldn't be an issue in long exposures.
2) Photon Shot Noise, due to the discrete nature of light (see below).
3) Dark current noise in the photodiode and amplifier (also described below).

Light appears to be continuous to us but with the photoelectric effect, light is captured in discrete packets called photons. As wikipedia will tell you, the energy of each individual photon is determined by the wavelength (colour) and the strength of the light is affected by the number of photons arriving per unit time.

Arriving photons are captured by each sensor pixel and (in principle) each photon arriving knocks an electron out of the semiconductor. It is attracted by an electric field that accelerates it toward a semiconductor junction, where its arrival will create a small electric current that can be amplified and digitised to indicate the light level on that pixel during the exposure. In reality not every photon creates an electron and the proportion that do is known as the quantum efficiency, for Sony cameras mostly between 0.3 (A58, A900) to 0.6 (A7s, A9).

Because photons are discrete there is inevitably some randomness in the arrival time of the photons, and this is effectively noise, as it is a random variation between adjacent pixels receiving the same nominal illumination. This noise is known as photon shot noise, and like most similar processes it obeys Poisson statistics for which the noise behaviour is known.

If you take a picture of a well-lit white sheet of paper with a bright exposure, each individual pixel in a camera sensor receives several thousand photons during the exposure interval, and produces a lot of electrons. The shot noise produced by (say) 2500 electrons is roughly the square root of the number of electrons, or about 50. So if one pixel receives 2500 electrons the next might get 2550 and the next 2450 if all are illuminated the same (we'll ignore pixel colours here). The variation between pixels noise is around 2% of the total and isn't very much certainly a tiny fraction of one EV.

But a dimly lit area of the sensor may get only a few photons and produce (say) 3 electrons. Adjacent pixels might then get anywhere between 1 and 5 electrons. The noise is much smaller in absolute amplitude than the bright-pixel case, but it is a much higher proportion of the actual signal at least 1EV. So the amount of noise, relative to the pixel brightness, is much greater in areas with low illumination.
Is this sensor noise? No. When there is no light there is no photon shot noise (or at least very very little, as photons can be created spontaneously at low levels due to 'zero-point energy').
Most 'sensor' noise is either due to thermal noise in the amplifiers or to dark current noise in the photocells. Sometimes an electron jumps out of the silicon, and is accelerated by the electric field, even without an incident photon, and this produces a false current and a false digital value. This is not commonly an issue except in very long exposures, but is made worse by small flaws in the silicon which can produce high dark current (hot pixels). Using long-exposure compensation can detect this and cancel it, but it can only cancel the average dark current, and the noise still gets through.

Anyway, the point of this is that in a well-designed sensor, and in exposures not long enough for hot pixels to show up, the dominant source of noise in the pixel is photon shot noise. But this only occurs in pixels actually receiving light. The digitised value of a fully dark pixel is very low.
So, in a real-world image taken at high ISO, the noise most obvious in the image is not in fully-black areas of the scene but in areas which are dark but not black, because these are the ones which receive few photons and produce high amounts of shot noise.

If you don't believe me, you can try it for yourself, as indicated below.
But if true it means that most of the comparisons of sensor noise performance are quite irrelevant to real photographic images, because sensor noise can only be found in areas which are black and therefore not seen and the areas which do show noise are not black, but are lit. In other words, the noise you see is due to the small number of photons arriving at the pixel, and not due to any deficiencies of the camera.

Noise levels go up in high-ISO pictures not because the camera or sensor is poor, but because the amount of light received is low. You may have seen images from low-light cameras with photomultiplier tubes, where every photon causes the emission of many thousands of electrons. But the images still look very noisy, because the noise is in the arrival of individual photons themselves, and is thus part of the light being received.

It is true that some sensors are more sensitive than others, and that some cameras give lower apparent noise than others at high-ISO. This is likely to be because (a) the actual sensitivity is different (eg higher QE, especially comparing A7/A9 generation cameras with the A900) and thus the light level being seen is different, and (b) because high-ISO noise reduction is different, often because it makes the grain size of the noise larger and more obvious.
But again this only affects the noise in black or very near-black areas, and the slightly brighter areas will have more noise photon shot noise.


Edited by Miranda F - 13 January 2018 at 21:48
A900, A58, 5d, Dynax 4, 5, 60, 500si/600si/700si/800si, various Sony & Minolta lenses, several Tamrons, lots of MF primes and *far* too many old film cameras . . .
Back to Top
Miranda F View Drop Down
Senior Member
Senior Member

Joined: 11 January 2014
Country: United Kingdom
Location: Bristol
Status: Offline
Posts: 2569
Direct Link To This Post Posted: 13 January 2018 at 21:30
The test.
Here is a colour test chart which contains graded bars of white, yellow, green, cyan, blue, magenta, and red, each with 16 levels from 100% to 0% brightness.

1. To check your camera, put the image full size on your computer monitor and take a picture of it with your camera using a convenient lens with a moderate focal length. To start with, turn all noise processing and DRO off and take pictures with base ISO, medium ISO, and high ISO. In each case, expose the image normally so the darkest row is fully black and the white edge is bright but not saturated. Turn room lights off if possible to avoid screen reflections.
You should defocus the lens slightly so that the pixels of your monitor are OOF and no moire is visible. You can zoom into particular colours, or the darker rows if you like.

2. Download the images you have taken to your monitor and look at them. If necessary, you can brighten or slew the tone curve on your photo app to see more detail in the dark areas.

3. Assess subjectively, where you think the most obvious and annoying noise can be seen. On my cameras, it is usually the red channel, usually somewhere near the bottom at ISO100/200 and around half way at ISO3200.

If the most annoying noise is in the bottom row, then sensor noise is your enemy. If it is anywhere else, sensor noise is irrelevant and photon shot noise is the problem.


Edited by Miranda F - 13 January 2018 at 21:56
A900, A58, 5d, Dynax 4, 5, 60, 500si/600si/700si/800si, various Sony & Minolta lenses, several Tamrons, lots of MF primes and *far* too many old film cameras . . .
Back to Top
Miranda F View Drop Down
Senior Member
Senior Member

Joined: 11 January 2014
Country: United Kingdom
Location: Bristol
Status: Offline
Posts: 2569
Direct Link To This Post Posted: 13 January 2018 at 21:44
The Results.
Here are images from the Sony A58, taken at 400 and 1600 ISO. They show noise in the richer-toned colours, especially red, but not at the bottom. Clearly photon shot noise.
   

If I use DRO (eg DR+ level 3) to bring up the dark areas, the noise there increases (particularly at high ISO) because I am amplifying a signal with few photons and more noise, not because the camera is bad.

Here are images from the A900, at ISO200 (base ISO) and 1600.
     

Again, the noise level in the black areas is invisible. Indeed, I have to deliberately lighten the black areas and increase contrast way up before noise becomes visible in the black areas, but it is still less than in the brighter areas:


Edited by Miranda F - 13 January 2018 at 21:53
A900, A58, 5d, Dynax 4, 5, 60, 500si/600si/700si/800si, various Sony & Minolta lenses, several Tamrons, lots of MF primes and *far* too many old film cameras . . .
Back to Top
sybersitizen View Drop Down
Senior Member
Senior Member

Joined: 04 August 2006
Country: United States
Location: California
Status: Offline
Posts: 13501
Direct Link To This Post Posted: 14 January 2018 at 00:28
One way to address what's going on is to use an audio system as an analogy. An audio purist wants a system with a very low noise floor and a very high saturation level. That means barely audible sounds will remain as pure (noiseless) as possible while extremely loud sounds will remain as unsaturated (undistorted) as possible. Photographic purists want something similar.

Originally posted by Miranda F Miranda F wrote:

It is true that some sensors are more sensitive than others, and that some cameras give lower apparent noise than others at high-ISO. This is likely to be because (a) the actual sensitivity is different (eg higher QE, especially comparing A7/A9 generation cameras with the A900) and thus the light level being seen is different ...

Yes, less sensitive (less efficient) sensors waste photons, but that's not all. You have to also consider the efficiency factor of all the supporting electronics in a camera. Inefficiency anywhere in the signal chain can add something to the noise floor, just as it can in an audio system.

... and (b) because high-ISO noise reduction is different, often because it makes the grain size of the noise larger and more obvious.

Most people who are picky about photographic noise and DR also hate in-camera high ISO noise reduction because it reduces detail, so they prefer to disable it whenever possible. (The same can be said of older audio noise reduction systems. Newer ones are almost certainly much better, but probably still short of perfect.)

But again this only affects the noise in black or very near-black areas, and the slightly brighter areas will have more noise photon shot noise.

In any imperfect recording system (the only kind we have?) there is always a noise floor, and with a camera the noise floor will generally be the same in every pixel (with some slight wiggle room for random variations) regardless of the actual light captured. If the signal is strong enough to fully mask the noise floor, we don't see the noise (nor do we hear it with audio), so we're happy. But if the signal is not strong enough to mask the noise floor, the accuracy of the resulting output is affected. In audio, this manifests as hearing something other than what was intended or what is desired during extremely soft passages. It means true silence is unavailable to the listener. In photography, this manifests as something other than accurate true black or near black being presented in the image.

Now it's true that despite all this, most photographic scenes and most audio programs are recorded and presented accurately enough for most people ... but picky viewers, like picky listeners, often want to adjust things to their own taste. They might want to 'bring out' minute detail in nearly dark areas - just as someone might want to 'bring out' minute detail in nearly silent audio passages - by applying some kind of gain to those almost signal-less chunks of data; and they want to encounter as little noise as possible when they do so. That's when better audio recording and better camera sensors (and better cameras overall) matter.

So the answer to the question Is sensor noise really that important? is: It depends on who you ask.
Back to Top
happyjack View Drop Down
Emeritus group
Emeritus group

Joined: 10 April 2007
Country: Australia
Location: Kingaroy
Status: Offline
Posts: 3186
Direct Link To This Post Posted: 14 January 2018 at 03:32
After reading that there is too much noise in my brain from over-saturation
Only one way to solve that: zzzzz
7D turned A77,77M2,MIN 50/1.4,MIN100/2.8macro,Sig10-20,F4-5.6,SAL70-300G,18-250,CZ16-80,SAL70-200/2.8,75-300,24-80/2.8SAM, TAm 150-600G2

My ZF
 



Back to Top
clk230 View Drop Down
Senior Member
Senior Member

Joined: 25 June 2014
Country: Australia
Location: Adelaide
Status: Offline
Posts: 192
Direct Link To This Post Posted: 14 January 2018 at 04:58
This is an interesting exercise Anthony. Here are the results from my A700. To my eye the noise is more prominent at a point several steps above the darkest patches.

A700 ISO100


A700 ISO1600


A700 ISO3200


Here's just the bottom half of the ISO1600 which has been lightened 3EV in DxO.
C & C always welcome
Back to Top
sybersitizen View Drop Down
Senior Member
Senior Member

Joined: 04 August 2006
Country: United States
Location: California
Status: Offline
Posts: 13501
Direct Link To This Post Posted: 14 January 2018 at 06:04
In all the photographed examples posted so far, the darkest three color bands are failing to register the tonal differences that are easily visible in the original. In the vernacular, the blacks are blocked in those photos. That's the kind of thing a camera with a lower noise floor and a wider dynamic range would handle better.
Back to Top
Miranda F View Drop Down
Senior Member
Senior Member

Joined: 11 January 2014
Country: United Kingdom
Location: Bristol
Status: Offline
Posts: 2569
Direct Link To This Post Posted: 14 January 2018 at 07:41
Originally posted by sybersitizen sybersitizen wrote:

In all the photographed examples posted so far, the darkest three color bands are failing to register the tonal differences that are easily visible in the original. In the vernacular, the blacks are blocked in those photos. That's the kind of thing a camera with a lower noise floor and a wider dynamic range would handle better.


Actually, on my monitor I can (just) see the difference between the lowest bands on all the images I posted.

Maybe your monitor is set to crush some of the black tones? But that doesn't really matter for this test so long as you (and your camera) can see the difference between *some* of the lower levels (and so long as the monitor brightness/contrast are not so badly misaligned as to crush everything except bright bits). In that respect, the behaviour of the monitor cancels out (which is why I used it for the source as well as the display) - you just have to compare the black ones on your image to those on the source.

I wouldn't disagree with your audio analogy, btw, except to remark that in audio the source noise is usually less than the noise in the transmission & reproduction equipment, whereas with DSLRs the source noise (photon noise) clearly is larger.
A900, A58, 5d, Dynax 4, 5, 60, 500si/600si/700si/800si, various Sony & Minolta lenses, several Tamrons, lots of MF primes and *far* too many old film cameras . . .
Back to Top
sybersitizen View Drop Down
Senior Member
Senior Member

Joined: 04 August 2006
Country: United States
Location: California
Status: Offline
Posts: 13501
Direct Link To This Post Posted: 14 January 2018 at 08:16
Originally posted by Miranda F Miranda F wrote:

Originally posted by sybersitizen sybersitizen wrote:

In all the photographed examples posted so far, the darkest three color bands are failing to register the tonal differences that are easily visible in the original. In the vernacular, the blacks are blocked in those photos. That's the kind of thing a camera with a lower noise floor and a wider dynamic range would handle better.

Actually, on my monitor I can (just) see the difference between the lowest bands on all the images I posted.

I didn't say no difference can be seen. I said the darkest three color bands are failing to register the tonal differences that are easily visible in the original. I don't think you can disagree with that. In fact, in most examples the darkest four are actually failing.

Maybe your monitor is set to crush some of the black tones?

No, my monitor is fine and it's well calibrated.

But that doesn't really matter for this test so long as you (and your camera) can see the difference between *some* of the lower levels (and so long as the monitor brightness/contrast are not so badly misaligned as to crush everything except bright bits).

It does matter to some photographers because some of those dark shadow tones would contain detail in real life scenes, and some photographers would insist on lifting shadows to see the details and they'd be disappointed to also see the resulting noise.

In that respect, the behaviour of the monitor cancels out (which is why I used it for the source as well as the display) - you just have to compare the black ones on your image to those on the source.

What cancels out? The point is that the source material is not being accurately recorded.

I wouldn't disagree with your audio analogy, btw, except to remark that in audio the source noise is usually less than the noise in the transmission & reproduction equipment, whereas with DSLRs the source noise (photon noise) clearly is larger.

Every electronic inefficiency in a camera can contribute to the cumulative noise. We cannot test a sensor by itself; it is part of a complex signal chain.
Back to Top
sybersitizen View Drop Down
Senior Member
Senior Member

Joined: 04 August 2006
Country: United States
Location: California
Status: Offline
Posts: 13501
Direct Link To This Post Posted: 14 January 2018 at 09:22
Here are my two tests using an A99 at ISO 100 and ISO 3200. These were RAW captures with no NR or other manipulations, and converted with the same neutral settings. I offer links to the full size images so the noise comparison can be clearly seen.

http://thesybersite.com/public/Dyxum-Test-Chart-ISO100.jpg

http://thesybersite.com/public/Dyxum-Test-Chart-ISO3200.jpg

On my system, the additional noise in the high ISO shot is readily visible throughout the entire image, in every column, and in nearly every level within a column. The degree of high ISO noise in the shadows and 'blacks' of course becomes even more pronounced when those tones are lifted. All this is exactly as expected.
Back to Top
pegelli View Drop Down
Admin Group
Admin Group
Dyxum Administrator

Joined: 02 June 2007
Country: Belgium
Location: Schilde
Status: Online
Posts: 23526
Direct Link To This Post Posted: 14 January 2018 at 10:03
Originally posted by Miranda F Miranda F wrote:


Noise levels go up in high-ISO pictures not because the camera or sensor is poor, but because the amount of light received is low. You may have seen images from low-light cameras with photomultiplier tubes, where every photon causes the emission of many thousands of electrons. But the images still look very noisy, because the noise is in the arrival of individual photons themselves, and is thus part of the light being received.
If this is true why do modern sensors produce images with much lower noise than older sensors (at low light amounts/high iso)?
Mind the bandwidth of others, don't link pictures larger then 1024 wide or 960 pix high, see here
Back to Top
Jonas A-R View Drop Down
Senior Member
Senior Member

Joined: 29 December 2007
Country: Denmark
Location: Denmark
Status: Offline
Posts: 1207
Direct Link To This Post Posted: 14 January 2018 at 12:29
Originally posted by pegelli pegelli wrote:

Originally posted by Miranda F Miranda F wrote:


Noise levels go up in high-ISO pictures not because the camera or sensor is poor, but because the amount of light received is low. You may have seen images from low-light cameras with photomultiplier tubes, where every photon causes the emission of many thousands of electrons. But the images still look very noisy, because the noise is in the arrival of individual photons themselves, and is thus part of the light being received.
If this is true why do modern sensors produce images with much lower noise than older sensors (at low light amounts/high iso)?


Because they are more (quantum) efficient and have less read noise.

That obviously implies that sensor noise indeed is important which is the truth of the matter. i think the only way to arrive at a different conclusion as above is if the black point is set inappropriately high thus hiding noise.
a7r a9 35/2.8 55/1.8 85/1.8 90/2.8 28-70/3.5-5.6 70-200/4 LA-EA3 LA-EA4 16/2.8 fish 20/2.8 24/2 35/1.4 85/1.4 135/1.8 70-400/4-5.6ii
Back to Top
sybersitizen View Drop Down
Senior Member
Senior Member

Joined: 04 August 2006
Country: United States
Location: California
Status: Offline
Posts: 13501
Direct Link To This Post Posted: 14 January 2018 at 17:18
I was thinking some more about the audio recording analogy. A microphone is a transducer that converts sound to electrical impulses, and a camera sensor is a transducer that converts light to electrical impulses. Recording any phenomenon that is too 'weak' for the designed efficiency of the transducer is going to lower the quality of the first stage of the signal chain, which is never a good thing. BTW, EVFs, LCD screens, and monitors are also transducers, for whatever that's worth. No major revelation, but it's the first time I ever considered the word transducer in reference to photography.

I assume there are specialty mics designed to be much more sensitive to sound than mainstream mics. Is that true? I'm not talking about mics that focus sound with a parabolic mirror or other mechanical aids, but mics that are inherently more efficient.

Are things now at a stage where camera sensors are nearly as sensitive as they can be? And why is 'the best' ISO typically 100 or something not far from it? Why is it not something much higher? Is it just because it would limit good (saturation-free) recordings to low light situations only, or is there something about silicon that determines that?
Back to Top
Dyxum main page >  Forum Home > Equipment forums > Camera Talk Page  123>

Forum Jump Forum Permissions View Drop Down



This page was generated in 0.092 seconds.

Monitor calibration strip

Dyxum.com - Home of the alpha system photographer

In memory of Cameron Hill - brettania

Find us on Google+

Feel free to contact us if needed.