Re: With 12/16bit RAW do you still need an ND Grad

John Sheehy <JPS@xxxxxxx> wrote:
Floyd Davidson <floyd@xxxxxxxxxx> wrote:

A 14 bit linear data file has a "useful dynamic range" of 11
fstops,

"Useful" is a subjective thing. It is dependent upon the use and the
user. To speak of "dynamic range" as a monolithic property of an imaging
system is folly, unless you give an objective criterion.

Basically true, except that "useful dynamic range" is merely a
way to force the use of a *valid* criteria as opposed to some
objective but virtually useless one.

It is a reality check to maintain perspective. (Something you and
Roger Clark both need a little more of.)

The most common
is maxsignal:noisefloor; that comes from audio, I think, where there is
no analog to shot noise; only to read noise.

Impulse noise (phase hits and amplitude hits) is a significant
characteristic of audio transmission systems.

And we have quantization distortion to deal with too.

Regardless, comparing "maxsignal:noisefloor" may or may not be
the most common, but it certainly is *not* the most useful with
audio work (nor with photography). A more comparable "audio"
measurement is the standard practice within the
telecommunications industry used to measure noise in digital
channels carrying digitized analog signal. It is essentially a
standardized SNR measurement. And that is of course exactly
what we are talking about here too, so there is a great deal of
similarity.

Typically what is measured on digital telecom circuits is called
"C-weighted Noise with Tone". A -10 dBm0 signal (which is to
say that it is 10 dB below the maximum desired signal, which
also happens to be about 4.1 dB below the absolute maximum,
making the -10 dBm0 about -14.1 dB from absolute maximum) is
applied at one end, and a notch filter is used at the distant
end to filter it out while the remaining noise in the channel is
then measured. A "C-weighted" filter at the measuring end
adjusts the frequency response to match the effects of noise
interference on a human ear.

Significant points to keep in mind are that a standard signal is
used, a standard frequency response (matched to the effects of
the noise measured) is used, and that quantization distortion is
included in the measure results. Such measurements are
comparable between different measuring devices and different
transmission facilities or configurations.

By one level of usefulness,
system A can have more useful DR, and by another, system B can have more
useful DR.

That is precisely what I'm trying to get across. We cannot just
go hunting for a set of definitions that provide the biggest
numbers, and then say that one system is better because its
biggest number is larger than another system's biggest number.

Comparing the 12-14 fstop dynamic range of analog output from an
electronic sensor is useful for comparing different sensors,
including film; but those numbers do *not* compare directly to
the dynamic range discussed when comparing 12 bit raw data files
with 8 bit gamma corrected JPEG formatted data if they are to be
used for printing or display on a computer monitor.

Apples and oranges.

This is because noise is not a simple function of signal; it
has multiple analog components, which can be totally unrelated to signal

The problem is that what we measure, and call "noise", may not
actually be a direct measure of the interference with the
signal. One noise can interfere more than another while our
methods for measuring noise will equate them, or ignore some to
emphasize others.

That is why a reality check is necessary. Call it "useful
dynamic range", because that will be a reminder that ignoring
the effects while measuring quantities does not provide valid
comparisons.

There are any number of different examples. With audio the
difference between interference by a single tone as compared to
an equal amount of power in a bandwidth limited gaussian
distributed noise is one example. Pattern noise in an image is
another example.

(sensor blackframe read noise, dark current noise, and ADC noise),
proportional to signal (what I call "scalar read noise"; present in my
Canon XTi but not noticable in my other cameras), and proportional to the
square root of signal (shot noise). Because of this complexity, it is
quite possible for one camera to have more stops between a S/N of 1:1 and
maximum signal than another, and the latter have more stops from a S/N of
10:1 to max than the first, at the same ISO. 1:1 is way down there in
the shadows for a Pentax K10D at ISO 100; up high for ISO 1600; 1:1 is up
high in most Nikons, especially at high ISOs. For Canon DSLRs, 1:1 is
pretty far down into the shadows, but not as far as at low ISOs or
especially the Pentax at ISO 100. 10:1 is much more related to shot
noise, in general, and has more to do with exposure and overall quantum
efficiency than anything else.

That is precisely why "useful dynamic range" is required to make
comparisons. The exact measurements of different types of noise
is _not_ an exact comparison between interference levels.

You, for example, have totally ignored quantization distortion,
which will, in many cases, totally eclipse all other noise
sources that are present in your 1:1 examples above.

When the dynamic range of the analog sensor
approaches 16 bits

16 bits is not a dynamic range;

Sigh. Don't play word games. I can use a lot more words to pin
that down very precisely, but it serves no purpose other than to
make long and boring articles that are not fun to read.

it is a potential *limiter* of dynamic
range,

So my above statement is precisely correct, and not overly
wordy.

but until analog noises get much lower than they are now, 16 bits
would do nothing much more than coax converters to work possibly with
more precision.

That is not true. Quantization noise from a 12 bit codec is
often greater than other noise sources, and would be reduced if
14 or 16 bit formats were used. The problem with that of course
is the required hardware will be slower, more expensive, and
will limit the number of images. Currently is appears that
almost all manufacturers of DSLRs have found the advantages of a
12 bit codec greater than the disadvantages; but that will of
course change in the relatively near future as speed and memory
capabilities of various components are improved.

In all the tests I've done, it is clear that the noise
in ADUs or DNs needs to be below 1.0; close to 0.5 before the bit depth
becomes a significant cause of noise and limiter of DR.

Are you even attempting to measure quantization distortion?

Here's 4 versions of the 14-bit ISO 100 output from the 1DmkIII that
Imaging Resource has; it's a 100% crop of RAW data with simple RGB
interpolation and white-balancing, of a shadowy area of a girl's hair,
pushed to about ISO 3200 for viewing purposes. The one marked 14-bit is
the original, and the lower bit depths are from quantizing the RAW data
before any other conversion steps, to 12, 11, and 10 bits. All images
are interpolated and white-balanced with 256 levels from the 14-bit RAW
represented with the full DR of photoshop's 15/16-bit mode (so none are
posterized any further by the processing). The blackframe read noise in
the resultant DNs is 4.88, 1.22, 0.61, and 0.3, respectively. As you can
see, there is almost no visible difference at all between the 4.88 (14-
bit) and the 1.22 (12-bit). The 0.61 (11-bit) starts to show some small
quantization noise, and the 0.3 (10-bit) shows it very clearly:

http://www.pbase.com/jps_photo/image/76001165/original

Reality check:

Linear data dynamic range
bit depth > 4 levels per fstop

10 7
11 8
12 9
13 10
14 11
15 12
16 13

If your results do not correspond to the above, with reference to
posterization, your methods/analysis are invalid.

--
Floyd L. Davidson http://www.apaflo.com/floyd_davidson
Ukpeagvik (Barrow, Alaska) floyd@xxxxxxxxxx
.

Relevant Pages

  • Re: phonograph to usp adapter
    ... mic impedance to get a flatter frequency response from the mic. ... faint electrical hum from my turntable is really mechanical noise. ... there is still the matter of lack of RIAA response curve. ... reaches 16 bits broadband dynamic range. ...
    (rec.audio.tech)
  • Re: To RAW or not to RAW?
    ... >> obviously visible in the lower brightness areas, ... Clearly noise (read noise, as well ... the dynamic range of the quantizer). ... >> the design teams for the 6 cameras you say would benefit from ...
    (rec.photo.digital.slr-systems)
  • Re: To RAW or not to RAW?
    ... http://www.clarkvision.com/imagedetail/digital.signal.to.noise shows sensor dynamic range and indeed some sensors do have more then 12 bits of dynamic range. ... out of 4 of the DSLR sensors have a sensor dynamic range above ... under table 5 it also says: "Signal-to-noise assumes photon noise limited. ... It has a read noise at ISO 100 of about 16 electrons, and a full well of about 44,200 electrons. ...
    (rec.photo.digital.slr-systems)
  • Re: To RAW or not to RAW?
    ... to 8-bit jpeg, and includes Poisson noise from photon counting, and for 3 types of cameras, from point and shoot to high end DSLR. ... So much for the supposed dynamic range of a JPEG format. ... at ISO 100 of about 16 electrons, ...
    (rec.photo.digital.slr-systems)
  • Re: Jitter measurement
    ... frequency is driven to ensure that at midnight the total number of cycles in the ... >Assuming you can make three of your PLL circuits, ... >the noise they introduce by comparing their outputs. ...
    (sci.electronics.design)