Is it true the rumoured D800s will have 1 stop iso improvement in the JPEGS ONLY, and not RAW?

Started May 30, 2014 | Discussions thread
noirdesir Forum Pro • Posts: 13,381
Re: Is it true the rumoured D800s will have 1 stop iso improvement in the JPEGS ONLY, and not RAW?

mlewan wrote:

noirdesir wrote:

mlewan wrote:

falconeyes wrote:

G Davidson wrote:

I have to admit to not knowing much about the physics involved, it just seems like way too limiting a statement

As some have said, I refer to physics with my statement.

To be more precise: There *is* room for a 1+ stop improvement in the future, but *not* within the limited scope of what Nikon/Sony can do for a hypothetical D800s sensor. So, there won't be that 1 stop improvement in raw. Period.

A period after a statement without any source whatsoever to back it up? It looks like a bunch of confused question marks to me.

What kind of source do you want?

Any kind of source. Absolutely any that does not sound like a tired parent telling their three year old "because I tell you so".

A link to physical research in the area would be fine.

A link to a wikipedia article would be nice.

A mention of a book which talks about the subject would be ok.

A link to a blog of some accepted expert on the subject would be ok.

A link to some techie magazine that writes about it would be ok.

Even a mention of a buddy down the pub who had met someone who had claimed this was the case would be better than the completely void and unverifiable "Period" above.

So, what will I find when I go through your posting history and check how many of your statements have been backed up by a source? I'm sure you might refer to DPreview, DxO, lens tests sites and so on but that only works for statements about a specific product not for more general ones.

Generally a source is accepted as backing up statements when one or both of the following things are true: (1) The source explains the issue in more detail and makes conclusions whose logic are at least partly understood or (2) the source is accepted as an authority, wherin (1) re-inforces (2). The (1) category is generally the better because you don't just have to believe somebody but can reach the conclusion yourself based on the presented data. However, even though people writing here "The quantum efficiency (QE) is already at 50%, more than 1 stop of improvement is thus physically impossible." would fall into the (1) category but is not recognised by you as such, probably because it doesn't start deep enough and is not understood because it relies on missing prior knowledge.

I can barrage you with links that explain the physics, the terminology and present measured or calculated results. However, a statement saying: "Moreover, no sensor can bring a 1 stop improvement in light sensitivity. Not anymore, we are already too close to physical limits now." will be harder to find because it essentially is taking several aspects together and putting one and one together which will seem to trivial for a lot of sources of the underlying physics. But I'll give it a start.

What is the source of noise in digital images?

The so-called photon shot noise and the read noise which is actually an umbrella term for all noise added by the sensor electronics between the generation of photo electrons in each pixel and having a value of it in digital form.

What sensor properties influence how much noise its images have?

Shot noise is reduced by increasing the fraction of the incoming photons that are registered (current sensors don't register photons individually but just the 'sum' of the photo electrons that are created when a photon 'successfully' hits the light-sensitive area, the photodiode, within a pixel, one photon creating one electron in that case, in the form of a charge or voltage). The efficiency of this is called quantum efficiency (QE) where 100% would mean every incoming photon creates a photoelectron. It being less than 100% meaning photons don't reach the light sensing area or some photo electrons get lost before their sum is 'counted' and converted into a digital value. Reducing the number of photons 'counted' increases the shot noise in the same way as siphoning off photons with ND filter or any other means. Obviously, doubling the QE allows to have the same shot noise with half as many photons reaching the sensor.

The other obvious path is to reduce the read noise (I hope I don't have to present a source for this statement). Of interest now is naturally which of these two main sources of image noise is the dominant one and under which situation.

How to express the amount of noise?

The normal way of expressing noise is standard deviation of the recorded signal for a uniform input. Shot noise is a variation of photon in time but since the photon stream hitting one pixel is independent of the photon stream hitting the next pixel (as the result of flipping a coin is independent of the previous flip) the variation in space of photon count, which translates into variation in brightness in the image and thus what we call noise, has the same statistical properties.

In a viewed image, the absolute brightness always varies (at most) between the maximum and minimum of the viewing media. What we care about is the relative standard variation, ie, the absolute variation divided by the average signal, for that viewing medium, not the absolute variations in light emitted or reflected from the medium. Thus what matters at pixel level is the relative standard variation in the number of electrons 'counted'. The inverse of the relative STD (deviation aka noise/ signal) is signal to noise ratio, SNR (signal/noise).

Now, how many electrons are typically counted in a pixel? That depends obviously on the size of pixel and how large the exposure is. Since this thread is mainly about the D800, let me quote some values for it. The best way to start is with the maximum number of electrons before the pixel becomes saturated. At base ISO (100) that is about 45 000. At higher ISO that number is capped not by the pixel properties but by the choice of not 'counting' numbers above a certain point (that limit can be imposed on the analogue side by amplifying the analogue signal and then running into the physical limit on the analogue to digital converter). If this treatment meant exactly the same for each ISO level, doubling the ISO should half that number. In reality, there is some deviation from that (not least since there are different ways to define ISO and the camera makers have a lot of leeway there). So, at ISO 3200, that number is about 1800 (instead of the 1400 of strict scaling). The minimum number of electrons would in principle be zero in total darkness if there were no noise added by the sensor.

This means six stops below pure white at ISO 3200 (which in most images at that ISO would likely be the dark shadows) would correspond to 66 electrons. The shot noise at that level is about 5 e- whereas the read noise is 3 e- (combined thus 0.12 STD). Thus even at this low level, the shot noise component is twice as large as the read noise component. As the absolute amount of read noise is about the same for all ISO values but the absolute amount of shot noise increases (by the square root of the signal increase), its relative contribution decreases as we move to lower ISO values.

Shot noise at at six stops below pure white was estimated from SNR at 18% at ISO 3200 (18% being about three stops below pure white), which is 31.3 dB, assuming it to be only shot noise (which it is not fully thus very slightly overestimating shot noise), the fact that shot noise decreases by 3 dB using DxO's definition of dB for every stop (as shot scales with the square of the absolute signal), thus ending up with an estimate of 22.3 dB for three stops below 18%. Converted into rel. standard deviation 0.077, multiplied with 66 e- getting 5 e-.

How much room for improvement is there?

As already said, to gain one stop, when only looking at the shot noise, we would need to keep the number of 'counted' photo electrons constant while halving the number of photons (one stop less light), in other words, doubling the QE. Current sensors have about 50% QE (the best ones a bit above) which means since we cannot go higher than 100%, we can reduce the shot noise component at most by one stop.

For the read noise the theoretical limit is zero. In essence doubling the QE would allow us to get the same 66 e- for half the light (one stop less) combined with zero read noise, we'd had only 5/66 or 0.076 as our rel. standard deviation, going down one more stop we get 0.11, ie, we got two stops and about the same noise. But that refers to the dark shadows at ISO 3200. If we were to look at the mid gray (18%) at ISO 3200, the shot noise would be clearly dominant and thus its reduction being the limit.

If you want some really informed discussion about sensors, have a look at Eric Fossum's posts here. He is one of the key people behind the inventions of the CMOS sensor and is still working in image sensor development. I don't think there are a much better qualifications to have on this subject.

P.S.: Note that all but one of my sources were Wikipedia or DxO on articles you could have found yourself (image noise, photo diode, definition of noise at DxO, quantum efficiency, shot noise). It's a variation of 'Let me Google that for you', with 'Let me look this up on Wikipedia for you'.

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow