Fast lenses, and High ISO

Started Jul 18, 2014 | Discussions thread
John Sheehy Forum Pro • Posts: 19,735
Re: Understanding ISO

Great Bustard wrote:

I've always wondered what the difference in QE is between the different color channels -- do you know?

If you shoot a gray card under mid-day sun, you get 1/2 stop less light captured in the blue pixels, and 1 stop less in the red ones, typically. And of course, if our ideal is a wavelength-sorting sensor rather than a CFA, only 1/4 of the pixels are that sensitive to red, and the other 3/4 are far less sensitive to red.

2-3 electrons is not a big issue at base ISO, with big pixels, but for small pixels and and high ISOs it is tremendous read noise.

This is why I said below:

For sure, the read noise represents a kind of floor beyond which the noise in the photo quickly diminishes.

Did you mean "above which"?

Well, yes, statistically, read noise diminishes greatly above the so-called noise floor at base ISO, but it can still be quite visible when it is correlated or has banding. Otherwise, ideal Poisson noise and Gaussian noise are indistinguishable as photon populations get above 12 (12 is the threshold that many Poisson noise generators use; above 12 mean photons, they typically use a Gaussian noise generator because it is faster). I've seen banding and blotching in highlight areas of base ISO when using a lot of sharpening, and contrast or saturation boosts.

And of course, real world read noise does not look like gaussian noise created by a computer; it is highly correlated, visually, even if not significantly so, as a statistic.

I'd be interested in more info on that. I've heard others say that read noise is pretty much gaussian (assuming we don't count banding as noise, since it is systematic).

This is one of the bad sides of knowledge, IMO. Lots of folks have learned about signal acquisition and processing in school or in their jobs, but a lot of times knowledge makes people too quick to deal with phenomena and close the case. Yes, the histogram of read noise is often close to pure Gaussian, but that tells you nothing about how the noise interferes with our perception of images. We don't see monolithic histograms of noise; we see noise particles of various shapes and sizes interfering with the discerning of detail in the subject, or creating distractions.

Imagine that you had a black frame from a camera with crazy low-frequency noise blotches and thick, heavy banding that caused different color casts in different parts of the frame. Now, imagine that the same exact pixels were in another frame, with the same exact histogram, but you randomly scrambled the pixels in the image so that all of the patterns of the original were gone. They have exactly the same histogram, possibly even near-perfectly Gaussian, but would they look anything alike? They would look different enough zoomed into 100% pixel view, but reduce them in size to the size of a business card, and the original is just a bunch of lumps and bands, while the scrambled/randomized one with exactly the same population of pixels and histograms is much closer to black. Histograms are only meaningful when the character of noise is already known and accounted for. This is why I mentioned to Jim Kasson that generating Gaussian noise in simulations is overly optimistic; it is more practical to use real-world black frames to simulate read noise.

That limit is much farther away, and is only limited in regard to how much we want to magnify the capture in display. Pure photon captures can work very well with very, very tiny signals if you do not need to display them large. They have no contrast issues in near-blacks, and no linearity problems down there, either.

The thing is, though, while lower read noise will certainly push the noise floor further out, the photo will still be rather noisy due to low signal. For example, while we may be able to shift the "noise floor cliff" from ISO 25600 to 102400 with lower read noise, we're never going to make ISO 25600 look as good as ISO 3200.

ISO 25600 will never have as many photons as ISO 3200, except maybe in the red or blue channel with an ideal wavelength-sorting filter system. The shadows of ISO 25600, however, would be extremely usable on a FF camera with no read noise.

Any read noise standard deviation higher than about 0.15 interferes with very high ISOs. 0.15 would be visible at very high ISOs, but since at 0.15 each photon count has its own virtually distinct histogram not overlapping the higher and lower photon counts' histograms, one could reduce all the values in the bell curve to a single value with a very high level of confidence. And of course, your digital gain must be high enough to render the individual bell-curves. I'd prefer a magical black-box photon counter and wavelength recorder, of course.

IMO, it is futile to try to predict the usability of extremely high ISOs with only photon noise based on monolithic SNR values as would exist in gray patches of various intensity, and assuming that photon noise is just like read noise but is proportional to the square root of the signal. The differences go far beyond that, because image subjects are not big gray squares. Images are modulating signals, where a thin white hair can exist against a black background, and with pure shot noise, there is no loud noise in the black part next to the white hair to obscure the hair. The white may not even be contiguous with a low signal level, but it still clearly suggests the line. I don't know where it is now, but a few years back I took a shot into an abandoned warehouse through a hole in a piece of wood in a window, with spiderwebs beyond the hole, and I decided when I got home to simulate, based on the camera's QE, what it would look like at ISO 6 million with only photon noise (of course, the simulation also included the noise from the actual capture, so it was worse than it would be with pure shot noise). Guess what; I could see the threads in the spiderwebs at 100%, against a dark background. At 6x4 inches or so on my monitor, you had to stare at the image to notice any noise, and it looked more like very fine sand art than it did a typical high-ISO digital capture.

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