@Jossie: by “post-proc gain” I mean afre’s “gain via post-processing”, ie setting a low ISO such as 100, then multiplying the image in the computer, after the work in the camera.
The other case is “gain via ISO”, where we set a higher ISO eg 6400 in the camera, so the camera does the multiplying by some secret process that may include some smoothing or other noise-reduction.
Yes theory is useful. And so is testing.
Adding additional analogue gain will add noise (and possibly non-linearity) to the signal. Digital gain will increase quantisation noise in the signal. A certain amount of noise on the input to the ADC is a good thing as it will dither away the effects of quantisation.
Which one is best will depend on the details of the signal processing hardware…
I believe most cameras end up doing a combination of both.
Assuming you keep same exposure for two photos of same scene , one taken at ISO 6400 and one taken at ISO 100.
Knowing that the analog amplification is done as far upstream as possible, the ISO 6400 image is theoretically less noisy than the digital amplified ISO 100 image.
In case the noise introduced by amplifier and downstream electronics (mainly ADC) is sufficiently small, it could be negligible versus the upstream noise. Then , the two images above (ISO6400 and ISO100X64) will present the same noise level.
I think In-camera digital gain (extended ISO) is of no value if you process raw file.
Is there any domain, where the digitization noise is comparable or larger than the shot noise?
See comparison above.
If the camera is well designed, the shot noise should always be the limiting factor.
Hermann-Josef
But the sensitivity of the system. It might be true that
But it is always wise to amplify early to not amplify later noise sources. If there are no highlights clipped, it is wise to amplify the sensor’s signal to best utilize the AD converter’s dynamic range, that means, ETTR. And reasonable ISO settings would do that.
If it is true that some cameras do not change this first amplification’s gain based on ISO setting, there may be a possibility to implemet ISO invariance in the sense that there would be no difference if I set ISO in-camera or boost exposure later. However, there are still some noise sources along the way, analog and digital ones, in camera and in the raw procesor, until the signal gets to the exposure setting in the latter. However it is implemented in-camera, I clearly see advantage in setting the ISO correct. The only way I see how this could be changed is some auto-gain in the camera. Maybe that’s how these ISO invariant cameras are implemented? One could easily check this with an underexposed raw file.
I was explaining @snibgo observation.
I suppose that for professionnal astronomy sensors, to fight thermal electronic noise, you can cool sensor and the electronic devices with some coolant (liquid helium for instance) and use special electonic , ADC and sensors designs…
In this case, it is possible that the only noise you have to take into account is the photon shot noise.
Your system is thus iso-less
It’s not at all the case for commercial camera (see also Emil Martinec paper). And in certain situations, camera electronics can be very hot thus generating a lot of thermal noise.
Thats surely mandatory for astronomy systems, but not for commercial cameras where the upstream (front end ) noise includes other noise sources.(see Emil martinec paper)
In this case your camera is ISO-less. That’s not always the case for commercial cameras at low ISO but becomes rather true beyond some ISO depending on the camera.
About ISO invariance see
This is actually funny. It is clearly visible that the image boosted from ISO 400 has much worse magenta-type noise as expected. Examine carefully the shadow below the painting at 7:03.
Yes, astronomical CCD cameras are cooled by liquid nitrogen (He is not necessary) to about 150K and the electronic parts are in most cases selected for minimum noise. However, the photon flux levels are orders of magnitude less than in photography. Exposure times (relevant for dark current and associated noise) are also on the order of up to an hour, instead of 1/100 of a second. If you look at Fig. 13 in Martinec’s work you see that at average illumination the shot noise in digital SLRs is, as I had written, the dominant part.
The discussion is going now in a somewhat different direction from the original topic. But still the bottom line here is that ISO does not change the sensitivity of the sensor but only the gain, i.e. the amplification factor of the voltage, delivered by the sensor.
Hermann-Josef
Unfortunately, my electronics knowledge has atrophied somewhat since I left university 30+ years ago and started writing software, so I’m really out of date on the specifics.
Others correct me if wrong, “digitization noise” is probably more specifically termed “quantization error”, where a specific analog value has to be shunted to the nearest digital quantity. N’est ce pas?
If so, then the domain of interest would be where the destination bit depth becomes “comparable” to the error magnitude characteristic of the shot noise… ??
Thinking out loud…
I think this is correct, it is a “rounding error”. So this may only be relevant for 8bit/pixel data and for those in the darker parts, I would conclude.
Hermann-Josef
That’s my surmise also, but I’m not sure how I’d quantify it.
Quantization error, aka rounding error, is discussed at http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/index.html#quanterror :
It is also relevant when deliberately underexposing, then multiplying in the computer (“post-proc gain”). In my post above, the ISO 100 photo is underexposed by 6 stops. The camera records 14 bits/channel/pixel, so the top 6 bits are zero, so the camera is using only the bottom 8 bits. The raw files are effectively 8-bit. Of course, they are linear.
This causes some noise. Worse, we can expect severe banding in the shadows. Is this banding worse in post-proc gain, or in ISO gain? I haven’t tested this.
The only shadow banding posterization I’ve experienced was due to big-integer → little-integer colorspace conversion to my display profile. Vexed me for a bit, as the output JPEGs always looked fine. Went away when I refactored my display color conversion pipeline to go straight from the internal float to the 8-bit integer for display.
Noise has been my challenge, just as you describe. My new camera has a highlight-weighted matrix metering mode, which apparently places the brightest parts at middle gray. I like that images exposed with that mode don’t have saturated channels, but if there’s a high dynamic range in the scene, those dark parts pulled up with tone curves can be noisy.
Next experiment is to shoot that metering mode wtih a +1EV exposure compensation. Some say I’ll blow out some images, but I’m not sure; in my last locomotive shoot I took pairs of exposures, one with regular matrix, quickly followed by one with HW-matrix. The EV differences of the pairs ranged from 2 to 5, but in all the HW images the bright parts were at least a full stop below saturation.
Edit: banding posterization, following @paulmiller… 
Actually, the noise helps to hide the posterisation in the same way that dithering on a ink jet print makes smooth looking tones from only 1-3 possible ink droplet sizes.
If the ISO gain in the camera is applied as an analogue gain stage before the DAC, there isn’t any extra posterisation at all - just more noise.
Edit 1: Aargh… ADC not DAC!
Edit 2: I mean posterisation, not banding. The dither effect only works if the noise is uncorrelated with the location on the sensor. If you have pattern noise (which shows up a streaks or stripes in the shadows) or hot pixels, digital scaling will make it more visible. Analog gain will probably make it more visible as well, depending on where in the system the pattern noise is being generated.
Sorry if I’ve caused confusion. Yes, representing an input range of values (eg photon counts) by a small range of numbers (typically integers, or some other quantization) reduces the number of distinct values, and we call this “posterization” (or perhaps “posterisation”).
However, the effect is typically not noticable except where the input range represents a gradient. A gradient input is transformed by posterization into visibly distinct bands. In linear encoding, the effect of posterization is strongest in the shadows. Shadow areas often have a strong gradient component, so we get “banded shadows”.
Yes, adding noise before the posterizing (or even after) will reduce or eliminate banding, thus concealing the posterization.
Hello,
just found an interesting feature of Nikon’s Z7 relevant to the discussion here on ISO settings. It uses two different read-out modes to reduce noise at higher ISO settings.
Hermann-Josef
I don’t understand what the author means. What is your interpretation on what is going on?
The sensor uses a ‘dual gain’ design that at higher ISO uses a lower-noise readout method in each pixel (but at the cost of capacity, hence it not being used in lower ISO modes). From ISO 400 upwards, once in this higher gain mode, there’s little benefit to increasing the ISO setting (you may as well keep the ISO at 400 then brighten your images when processing the Raws).
@afre
I do not know the electronic details. But the electronic noise does depend on the way you read-out the CCD. E.g. a fast read out can produce more noise. Nikon switches the read out mode above a certain ISO setting. What this exactly means is not specified. Why this reduces the capacity I do not understand. The capacity of a CCD is a hardware property (full-well capacity, i.e. the maximum charge a pixel can store). What is changed by increasing the gain is the numerical range that can be accommodated in e.g. 16bit unsigned integer representation. With higher gain you run into numerical saturation at lower charges / voltages.
In any case my conclusion would be that post-processing amplification of a low ISO image (below 400) will be noisier than using a higher ISO setting. Above a setting of 400, as is stated in the note, there should be no difference.
Hermann-Josef
I was also interested in the subject, but more in ‘what does this mean to me and is it useful to know about it?’ than in ‘what happens in the sensor?’.
After reading some of the articles in internet, i reached these conclusions:
What is ‘useful’ to know about it (for my X-T2 which fails in the ‘dual gain’ category):
To make it short, I somehow decided that in general I will not see enough differences in ‘normal’ daily situations to bother to much about it. Just expose with whatever ISO fits the situation I need and forget about it. However, when doing night photos, especially with stars, if preserving the stars color information is important, I will expose at the minimum ‘high gain’ ISO (in my case ISO800) and fix the luminosity in post processing being careful to not blow away the highlights.