ISO Invariance

[Kayaker72]

Hi everyone....I have been mulling over some tech stuff for actually about a year+ (quantum efficiencies, differences in sensors, how much head room really exists over my Canon 5DIV, etc).

But as I am working on my thoughts which include how ISO works, Tony Northup came along and stirred up the "ISO Invariance" pot. I am not sure he even knew what ISO Invariance was when he did his first video, as he just said "ISO is Fake." Fstoppers then did their own videos, and everyone seemed to finally notice iso invariance even though people have been talking about it for 2-3 years and I think DPReview has had a test for it for ~2 years.

If you are not familiar, ISO Invariance in short is the fact with modern cameras adding "gain" (increasing exposure, etc) in post processing is almost identical to increasing ISO while shooting. And if you watch Tony's or Fstoppers videos, you'll see them underexpose images by 4-5 stops, lift them in post, and have the noise be almost identical to a shot that was properly exposed in camera. Note, this is really about under exposing, as a blown highlight cannot be recovered and, not mentioned in the videos, there are still numerous benefits to properly exposing such as working with flashes and cutting down post processing, and, in my experience, I find some images are just best when dialed in in-camera (colors, feel, etc).

But that said, the natural conclusion to ISO invariance is that you can under expose certain modern cameras with little penalty, and I do use this in my photography and this is one of the reasons I bought the 5DIV (that and f/8 AF points, etc). For high contrast scenes with a bright subject (loons, sunrises, etc) I often expose for the brighter part of the subject, underexpose the shadows, and bring them up in post. While ETTR still exists, it is less important (IMO) than it was 5-10 years ago.

But...all of this messed with my understanding of how ISO actually physically works. It is a bit funny that one thread is talking about old active users, as I am pretty sure my understanding of ISO and how it works came from Daniel Browning years ago. Maybe Neuro, or a combination.

My current understanding, camera sensors are charged diodes, and there are these basic steps:

1. The camera diode/sensor is charged
2. The charged diode releases electrons when hit by light
3. The released electrons are then captured by the pixel
   (electrons captured per photon of light is quantum efficiency)
4. In a CMOS sensor, the electrons are read by a transistor at the pixel level which creates an analog signal
5. Gain can be added to an analog signal
6. The analog signal is then converted to a digital signal
7. The digital signal is recorded as part of the RAW file.
8. Gain can be added in post processing

Some key things missing in the above that is relevant for some sensor discussions includes how the analog signal is read (in rows = rolling shutter; all at once = global shutter with obvious implications to on sensor circuitry/bandwidth) and where the analog signal is converted to digital (off sensor is the old method with more noise while on sensor has less noise as the signal does not have to travel as far).

So, with the above description, the question is, what is "ISO"?

Based on the TN and Fstopper videos and some online searches, what is being implied is that in camera ISO is simply an amplification to the analog signal in step 5. Whereas adding gain in post processing is simply adding to the digital signal in step 8. But Steps 1-4 never change while taking a picture.

Under this thinking, ISO invariance exists now because read/dark noise decreased to the point of being negligible when moving the analog/digital conversion on chip. In other words, since the digital signal is essentially as clean as the analog signal, it no longer matters very much where you add gain: in camera or in post processing.

Seems reasonable, I can find charts where that seems to make sense (DR/noise is linear with increasing ISO, etc).

But, it flies in the face of my understanding of what ISO is and I can also find charts where DR/noise are not linear with increasing ISO.

Quickly, my understanding of ISO was two things:

1. The camera diode/sensor received different charges at certain key ISOs, so the charge in step 1 changes at certain ISOs
2. Then the camera would add gain or subtract gain from the analog signal in this push/pull method that created this wavy DR/Noise graphs (step 5). Even today with a modern sensor like the A7RIII, there are two points from which it is linear, not one, implying two "base" ISOs.

But, I can't find anyone referencing variations in diode/sensor charge as part of ISO (variations to Step 1). It is all analog signal gain, or step 5 above.

So, after all that, the question is, does anyone know or know of a good reference for what ISO actually is? BTW, feel free to point out inaccuracies in the above. I am a Civil/Environmental Engineer, not an Electrical.

Thanks,
Brant

BTW, if you want examples:

Wavy results which to me imply either a variation in charge on the diode (step 1), or I actually see reference to a "dual gain" as explanation for the wavy behavior (so, a double step 5? or gain to analog and then gain to the digital signal?)

The blue (5DIII) line shows the wavy behavior while the black (5DIV) sensor having a little, but a lot less waviness and much straighter (iso invariant).

Then Sony sensors have straighter lines (which I can see being gain added at a step 5), but the A7RIII this one drop (which I suspect is something else and have previously thought it was a charge difference at step 1):

If interested, as Nikon buys their sensors, the behavior of their sensors varies with the source. But you can see wavy behavior, or straight line with one drop.

[Jonathan Huyer]

Interesting review... I'm fascinated by this stuff although most of it is way over my head (mechanical engineer, barely passed my mandatory electrical course 30+ years ago). I'd like to know where noise comes into play in all this. What is the intrinsic flaw (if that's the right word) in the electronics that creates noise? Do the diodes become less accurate as charge increases with ISO?

[HDNitehawk]

But that said, the natural conclusion to ISO invariance is that you can under expose certain modern cameras with little penalty, and I do use this in my photography and this is one of the reasons I bought the 5DIV (that and f/8 AF points, etc).

So wouldn't this make sense?

If your camera is set on the exact same shutter speed and aperture the sensor gathers the exact same amount of data regardless of the ISO setting.
If you adjust your picture to the proper exposure in post the data was always there, you would only be changing the point in the process it is done.

1. The camera diode/sensor received different charges at certain key ISOs, so the charge in step 1 changes at certain ISOs

I have never heard that one, why would it?

I do remember old conversations or debates about noise. A certain amount of noise is introduced at the sensor, additional (spatial) noise when you increase gain.
So perhaps if someone could explain how increasing the gain works, does it change the signal to noise ratio?

Engineers:
https://www.sis.se/api/document/preview/907324/
https://webstore.ansi.org/Standards/...SAAEgLtPPD_BwE

[Kayaker72]

Interesting review... I'm fascinated by this stuff although most of it is way over my head (mechanical engineer, barely passed my mandatory electrical course 30+ years ago). I'd like to know where noise comes into play in all this. What is the intrinsic flaw (if that's the right word) in the electronics that creates noise? Do the diodes become less accurate as charge increases with ISO?

Yeah....this Civil Engineer is breathing through a straw on a lot of this.

First, I am becoming convinced that I either misunderstood or whoever suggested ISO was based on varied diode charge was wrong. To me, it does make sense as increase charge would be more sensitive (fewer photons in low light but more electrons to get bumped off the sensor), but also more potential for noise (random electrons getting bumped off, more heat, etc). But, I even came across this reference where myth #1 on clarkvision states that "ISO changes sensitivity" is false. Then, i

n all the articles I have read that actually look this granular (which actually isn't many), I only see references to ISO being adding gain to the analog signal and that the diode/sensor to pixel/photon wells are the same for every ISO setting (i.e. base ISO). So, as the analog digital conversion has become more efficient (granted, only for some sensors), you get ISO Invariance.

Overall, it sounds like the "myth" of changing sensor sensitivity has been out there, I picked it up someplace.

It looks like the article I like for noise is no longer on the web (at least my link doesn't work). But in terms of understanding what is happening on the sensor level, this video is great (also talks about film, which is fun):

https://www.youtube.com/watch?v=MytCfECfqWc

Then I also like this article:

http://dpanswers.com/content/tech_iso.php (old...it is actually one of my first saved links for photography)

I came across this article while checking my understanding of ISO and ISO-Invariance. But it has a nice discussion of noise part of the way down:
https://photographylife.com/iso-invariance-explained

But, generally, there is noise associated with the variation of light (i.e. shot noise) that has nothing to do with our cameras, but then there is noise introduced during exposure, with inaccuracies in the voltage readout, while the analog signal is transferred (this is what plagued/plagues Canon), and then during the analog to digital conversion (again, any inaccuracies, electrons being added subtracted, etc). Overall, anything within the camera is generally termed "read" noise.

So wouldn't this make sense?

If your camera is set on the exact same shutter speed and aperture the sensor gathers the exact same amount of data regardless of the ISO setting.
If you adjust your picture to the proper exposure in post the data was always there, you would only be changing the point in the process it is done.

Yes, this makes sense for a truly ISO invariant system. However, in older systems and some modern systems, they are not iso invariant. For example, with ISO being added to the analog signal, all noise that happens while generating the analog signal is also increased. But all noise added while transferring the analog signal (in some cases, off chip) and while converting the analog to a digital signal is not amplified. Thus, with the exact same exposure settings you would get less noise in your RAW file if proper exposure was ISO 800 by setting ISO in camera to ISO 800 (adding gain to the analog signal) rather than shooting at ISO 100 (base ISO for most cameras) and adding 3 stops of gain post digitization/creation of the raw file. ISO invariance, the noise added after creating of the analog signal is negligible, so you could shoot ISO 100 and increase in post and have basically the same image. This is essentially what TN and FStoppers did.

For years, the knock on Canon sensors was that their analog to digital conversion occurred off chip, thus there was significant noise added while transferring the analog signal off chip. Also, I've seen write ups that part of the "1D" treatment, was multiple analog-digital converters, thus faster, less heat generated and less noise. But, starting with the 80D and including the 5DIV/1DXII/M6/M5/M50/R (maybe others), my understanding is that these cameras have on chip AD conversion, thus less transfer and less noise. These cameras, when you look at their noise/DR results, are pretty linear. In contrast, the 6DII seems to still be off-chip conversion. Still a great camera, but this is why some on the internet bash the 6DII.

I have never heard that one, why would it?

Yeah, I am pretty certain I picked up the "myth" somewhere. But, I picked it up, it made sense to me, and I saw data that fit the model (wavy noise results with varied ISO/quick change in noise results). But now, I can't find a single reference supporting this idea and have found several that seem to counter it.

So, I have areas of expertise, this is not one of them. It is always good to have some fluidity in your thoughts and opinions. Here, I am moving off this idea that different charges can be applied to a sensor and that contributes to ISO.

To explain the data, there has to be more to ISO than just gain added in one step to the analog signal. "Photons to photos" references dual gain to explain the wavy feature. Anyway, I'll post if I ever figure it out.

[Karsaa]

I actually did some testing with this invariance with my 5dIV.

What i did i took first picture of northern lights with settings like 4s iso 6400 f/1.8 then changed the settings to 4s iso400 and f/1.8. ISO 400 because apparently 400->5DIV is iso invarient. Then what i did on LR i boosted exposure that +4 and yes, the noise was close the same as taken with 6400. The dark areas on the landscape area do get some stripes what you get when boosting shadows too much etc. but the basics were pretty much the same.

I am wierdo when it comes to iso values when shooting northern lights that i don't use the so called base isos much i tend to use 2000/4000/5000 mostly when takin pictures. But what i need to test out still is that if taking the images like iso 800/1600 like i would do them with 4000/5000 and then do some +2-3 exposure upping, what kind of quoality i get there instead taking out some exposure from my bit over exposured ones.

[Kayaker72]

Hey Kari,

It might not be totally analogous, but what you are describing is similar to discussions I've seen for astro work. Of course, astro work could be stacking images or single exposures, but there is a lot of good discussion.
http://dslr-astrophotography.com/iso-values-canon-cameras/
https://www.diyphotography.net/find-best-iso-astrophotography-dynamic-range-noise/


Both those links get into the whole ISO-variant vs ISO-Invariant discussion. The first links to a pretty good discussion of ISO that pretty emphatically states that ISO is just Gain/amplification to the analogue signal:
http://dslr-astrophotography.com/iso-dslr-astrophotography/

I am curious where you end up as we are both shooting 5DIVs. I do not often shoot nightscapes, but would love to shoot more, so anything you can pass along would be appreciated.

For what it is worth, and this does not seem updated, but a few years back, over at Clarkvision they settled on the 7DII for their low light work:

http://clarkvision.com/articles/characteristics-of-best-cameras-and-lenses-for-nightscape-astro-photography/

Basically, very little dark current noise at given temperatures.


Brant

[DavidEccleston]

I think maybe the steps you list for the sensor are for CCD, not CMOS sensors. Chuck says "I cannot speak for other manufacturers' products, but EOS Digital SLRs with Canon CMOS image sensors do not carry electrostatic charges at any time"

[NFLD Stephen]

Seems like a lot of engineers on here...wonder if that says something about either engineers or photographers (or both)...

I read the same article Brant referenced from photographylife. That seems to make some sense to me, but I'm left with the question as to how some cameras and/or manufacturers can have such a low (practically zero) back end read noise while other cameras/manufacturers seem to have issues here. Surely it can't be that simple or beneficial or this would be something they would all be doing (or strongly trying to do).

Stephen

[Karsaa]

My wild quess on the read noise is what is the thing manufacturers want to favor. Apparently canon's idea is to have nice color space etc. which maybe then leads to issues in other areas. Not an engineer, but for me it looks like that the choice what is wanted from body is the somehow related to this.

[Kayaker72]

I think maybe the steps you list for the sensor are for CCD, not CMOS sensors. Chuck says "I cannot speak for other manufacturers' products, but EOS Digital SLRs with Canon CMOS image sensors do not carry electrostatic charges at any time"

Thanks. I just did a few searches to see if CCDs were handled differently from CMOS. It appears that the gain on a CCD is still to the analog or digital signals, same as a CMOS sensor. The more I look, the more references I find that there is just a myth out there that sensors have varied sensitivity. This is wrong. ISO should not be part of the "exposure triangle." Actually, there is no triangle. Rather, you have the light that hits the sensor controlled by aperture and shutter speed. Then gain can be added in camera (ISO) or in post processing (playing with exposure). In camera gain is usually added to the analog signal and post-processing is, of course, to the digital signal.

I believe chuck was talking about surface electrostatic charges that are rumored to attract particles. If he is right, this is another myth that is out there. I've always heard that sensors have an electrostatic surface charge. huh...well, this is what I like about being active on forums, I am still learning.

While I've seen this discussed several times, this video does a very nice job of laying out what a photodiode is, how it works, and the differences between CCD and CMOS:
https://www.youtube.com/watch?v=MytCfECfqWc

How the silicon sensor works, a diode, and the photoelectric effect are laid out from 2:50 to 5:30. A CMOS sensor is laid out at 9:04. Also, why CCDs are not great for modern cameras is laid out well in discussing the cons for CCDs starting at 8:10.

Seems like a lot of engineers on here...wonder if that says something about either engineers or photographers (or both)...

I read the same article Brant referenced from photographylife. That seems to make some sense to me, but I'm left with the question as to how some cameras and/or manufacturers can have such a low (practically zero) back end read noise while other cameras/manufacturers seem to have issues here. Surely it can't be that simple or beneficial or this would be something they would all be doing (or strongly trying to do).

Stephen

I've always been impressed by that correlation....lots of engineers (and some scientists) are hobbyist photographers.

For years the knock on Canon was too much background read/dark current/etc noise. This was largely attributed to having off sensor analog to digital conversion (ADC). Thus, the voltage for each pixel for each image had to migrate from the chip some distance before being converted to a digital signal. The longer the distance transferred, the more potential for slight changes to the voltage, which is noise. This affected blacks/shadows more simply because their voltage is so much lower than higher luminance.

So, if you have a signal of say 7 mV, and +/- 2 mV is added to it as noise during the transfer, that becomes an analog signal of 5 to 9. If you scale up after the ADC, say 3 stops (so 8x), that 5 to 9 luminance becomes 40 to 72. Had that same signal been set at ISO 800, so gain added before the off-chip transfer, then that 7 x 8 = 56 mV, then the +/- 2 mV noise is added and you end up with 54-58 mV range for the analog signal as it hits the ADC.

Sony gained a lot of notoriety for shortening this transfer distance and putting the ADC on the chip as the sensor. Canon has done that with some cameras (5DIV, 1DXII, 80D, etc) but only recently, and they still do not do that with all cameras.

The other primary difference that I've heard thrown around (but do not understand in its entirety) is chip architecture. Canon is somewhat unique in that they make their own camera sensors. Nikon buys theirs. Sony also makes chips and really got into making chips for camera phones.

This article is from 2012, so it is dated, but at the time Canon was still using a chip fabrication based on 0.5 microns (my understanding is this is the substructure that the chip is built upon). Sony, even in 2012, had 0.18 micron and 0.25 micron substructures. Again, Civil Engineer here, but smaller process generation allows for the construction of smaller, more efficient (less noise) systems on the chip level. It also seems that the smaller substructure may be important for moving the ADC onto the chip.

My understanding is that, again, 5DIV, 80D, 1DX II, chips are all built using a smaller process generation. I want to say I remember 0.25 micron, but I can't find that link to confirm, so maybe it is 0.18 micron that was referenced in the article. But, the fact that Canon has an existing "mature" 0.5 micron process generation system, they still occasionally pump out sensors with "old" technologies (hello 6DII) that it is inferred are based on the mature 0.5 micron fabrication (I've never seen this proven, but generally assumed).

So, there are a lot of differences between manufacturers. Patents likely ensure that will continue. Personally, while many focus on the above, Canon does many things extremely well and, as a photographer, I really have no issues with the images coming off my 5DIV and I had very few issues with my 5DIII and its "older" sensor fabrication. In fact, many of my all time favorite images were taken with the 5DIII. While I love the technical details, photography is still about capturing moments, scenes, and light.