Is the EPSON 4990 really 16 bit?

[Kennedy McEwen]

SNIP

Mostly because it depends on the subsequent workflow and also because
of limitation in out eyesight. There is also a difference between
"first generation" image data, with little if any processing done to it
except for gamma adjustment, and film+scan which I consider "second
generation" image data.

However, second generation film scans are intrinsically noisy due to the
grain that is present and should therefore be more immune to the
posterisation effects of a low bit quantisation. The opposite appears
to be true - film scans really need 14-16-bits to avoid posterisation
and even higher to reliably reproduce the density range of slides.

Any cumulative rounding to integer errors (e.g. due to post-processing
steps) will increase the chance of posterization becoming visibile.

<http://www.xs4all.nl/~bvdwolf/main/downloads/8vs16-bpch processing.png>
is a small crop from a digicam which, due to the low noise, is extra
sensitive to posterization.

Worse than that, I have some shots of the moon taken on my 5D with a
Tamron 500mm mirror lens and 2x extender. Even without *any* post
processing the lunar landscape appears disappointingly posterised on
close inspection. Even captured raw and processed off-line, which
should have more accuracy. For some reason I have yet to fathom, this
problem has only arisen on these images. I have taken dozens of night
shots of other, albeit more colourful, subjects and no sign of this
limitation has occurred. In fact, the density of the blacks in night
shots from the digicam is exceptional.

 

[Don]

The Bayer filter really only affects the resolution of the sensor and,
as Bart has mentioned, just takes advantage of a limitation of our
visual system that has been exploited since at least 1953 with the
invention of NTSC colour TV encoding, and possibly before that. It
shouldn't have any effect on the quantisation.

My thinking was that due to interpolation inherent in digicams (as the
4 Bayer pixels are converted into a single one) I expected
posterization to be masked to some extent because the 3 channels do
not correspond the same point in the real world. There's no such
interpolation when scanning film and the image is in *that* sense
"pure". However, a scan is second generation, there's grain, etc.

Because of all of that my instinct was that comparing such two
different paradigms solely on bit depth only tells half the story i.e.
just because 12-bit digicams produce an appearance of pleasing results
I didn't see that translate directly into abandoning 16-bit depth when
scanning because of all of these other aspects.

So the 12-bit linear encoding of the CCD output would result in visible
posterisation of the shadows from digicams if they encoded the gamma
directly as per sRGB or other working spaces require. I suspect, that
they deviate from that, thus avoiding the problem but rendering the
shadows somewhat darker than they should be. Of course, with slide
scans there is a readily available reference for comparison, the
original, which would prevent such fudges.

Exactly! That's why I was surprised that you seem to imply that 12-bit
dSLRs produce better results than 16-bit scans but I see now that was
not quite what you meant.

I see from your reply to Bart that you were really talking about
*appearance* of digicam output and that I understand and agree with
from first hand experience. So, never mind...

Don.

 

[Don]

The practical success of Bayer based CCD and
CMOS imaging rather gives the lie to the need
for 16-bit per pixel color depth.

Well, not quite because that's not comparing "like with like" which
was the gist of my message. There are far too many inherent
differences between the two paradigms to make such a direct comparison
and then draw conclusions based on such a limited comparison (bit
depth only) without taking into account all of the other aspects.

Think about it: your LS-50 or my LS-8000
capture 20 million *real* (non-interpolated)
RGB triplets from 35 mm color film.

My Canon 10D captures a mere 6 million
*interpolated* RGB triplets. And yet,
images from the 10D often produce the
better print.

Yes, but that's only appearance. If we reduce the resolution of a
scan, apply interpolation, etc. even in spite of grain and other
drawbacks of scanning such a "processed" scan will *appear* much
closer to a digicam shot.

I mean, a few years back shooting my slides with a Nikon CoolPix 950
produced "better" results than scans of the same slides with an LS-30.
And that's even if we ignore the cast problem.

More accuracy does not always produce better appearance. For example,
just look at the whole concept of anti-aliasing or "pro-blurring" as I
call it. :-/ "Fuzzifying" a diagonal line "hides" the jaggedness. To
most people that "looks better" but we know it's actually worse.

The same goes for scanning. For example, my LS-50 produces scans with
"pepper spots" but my LS-30 never had any problems with pepper spots
because it just couldn't resolve to such level of detail. But I would
not therefore call my LS-30 "better" than my LS-50 (in absolute terms)
even though (with respect to pepper spots) the LS-30 scans may "appear
better" at first glance.

Don.

 

[Kennedy McEwen]

On Sun, 29 Jan 2006 01:57:19 +0000, Kennedy McEwen

My thinking was that due to interpolation inherent in digicams (as the
4 Bayer pixels are converted into a single one) I expected
posterization to be masked to some extent because the 3 channels do
not correspond the same point in the real world.

It shouldn't affect the posterisation, which is easiest to see in slowly
varying densities. So the fact that the interpolation source pixels
don't correspond in position to the final pixel shouldn't make any
difference. So all that really matters is that you have two 12-bit
samples from which you are computing a third. The highest number of
equivalent bits that can exist in that interpolated result is 13, but
would usually be closer to the 12-bit original with a smarter
interpolation which biased on source pixel over the other depending on
the changes in adjacent pixels.

There's no such
interpolation when scanning film and the image is in *that* sense
"pure". However, a scan is second generation, there's grain, etc.

So reducing a scan to 12-bit linear, making colour separations, reducing
the resolution further to match the Bayer filter mask, then
interpolating back up to the original and recombining the colour
separations will make a non-interpolated scan look as posterisation free
as a 16-bit digicam output? I don't think so, but you are free to try.
;-)

 

[rafe b]

Yes, but that's only appearance.

What else matters?

If we reduce the resolution of a
scan, apply interpolation, etc. even in spite of grain and other
drawbacks of scanning such a "processed" scan will *appear* much
closer to a digicam shot.

Yes and no. I can downsample scanned 6x7 cm negatives
from 90 megapixels to 8 megapixels, and they're *still*
noisier than a capture from a Canon 10D or 20D.

More accuracy does not always produce better appearance. For example,
just look at the whole concept of anti-aliasing or "pro-blurring" as I
call it. :-/ "Fuzzifying" a diagonal line "hides" the jaggedness. To
most people that "looks better" but we know it's actually worse.

It sounds to me like you don't quite understand
what anti-aliasing is about.

You can "skimp" on anti-aliasing, but there's
generally a price to be paid -- as owners of the
Canon 5D are beginning to report.


rafe b
www.terrapinphoto.com

 

[Evo2Me]

Yes and no. I can downsample scanned 6x7 cm negatives
from 90 megapixels to 8 megapixels, and they're *still*
noisier than a capture from a Canon 10D or 20D.

1. Downsampling degrades an image as much as upsampling.
2. Since there is no difference to an algorithm or machine between
noise, grain and wanted information, all processing is applied equally
to all of them; they are all affected.
3. Are you really seeing noise (electronically generated) or could it
be the grain captured by the scanner?
4. Digital cameras produce smoother images than film due to the fact
that every pixel is created equal while film grain is decidedly
non-uniform.

 

[Don]

What else matters?

What's really in the file. Often times optical illusions make things
appear "better" than they really are.

However, if raw data is of superior quality one can always easily
achieve appearance equal to that of an "optical illusion". Indeed,
starting with superior data guarantees superior results (assuming a
competent operator) not just appearance of superior results.

But it all boils down to how discerning a user is and/or how much work
they are prepared to put into it all.

Yes and no. I can downsample scanned 6x7 cm negatives
from 90 megapixels to 8 megapixels, and they're *still*
noisier than a capture from a Canon 10D or 20D.

Because we are not comparing like with like (a scan is second
generation, for starters). A digicam picture may *appear* "better" at
first blush (e.g. no grain) but on closer inspection a scanned image
will be sharper (i.e. no Bayer interpolation) albeit that may only
amplify the grain, etc. Two different paradigms.

Above, I was just illustrating that appearance (i.e. subjective
perception) can be manipulated often making inferior data appear to
look "better".

Which again, comes down to user (personal) preference. At one extreme,
some are quite happy simply with what "looks better" while others
wring their hands and agonize over every pixel.

It sounds to me like you don't quite understand
what anti-aliasing is about.

What did I miss? In the above jagged line example, every time
so-called anti-aliasing is employed the data is degraded for the sake
of an optical illusion - which, BTW, I personally actually find worse!

You can "skimp" on anti-aliasing, but there's
generally a price to be paid -- as owners of the
Canon 5D are beginning to report.

That's different because it deals with frequencies in the sampled data
which do not exist in the analog original. Therefore, by judicious
application of filters one eliminates analog frequencies which
generate this "ghost" data in the sample.

In the above case of jagged lines, however, data is "invented" (to try
and "compensate" for lack of resolution) resulting in a "fuzzy" image
which is a concept I just deplore.

It's Zaphod Beeblebrox (sp?) style of "fix". ;o)

Any Hitchhiker's Guide fans out there?

Don.

 

[Don]

So reducing a scan to 12-bit linear, making colour separations, reducing
the resolution further to match the Bayer filter mask, then
interpolating back up to the original and recombining the colour
separations will make a non-interpolated scan look as posterisation free
as a 16-bit digicam output? I don't think so, but you are free to try.
;-)

Hey, I tried crazier things than that, remember! ;o)

But seriously, of course there is more to it than that because we are
not really comparing like with like.

Don.

 

[Marjolein Katsma]

Evo2Me ([email protected]) wrote in 4ax.com:

3. Are you really seeing noise (electronically generated) or could it
be the grain captured by the scanner?

How does one tell the difference? What if there's both a lot of grain
*and* noise - where does one stop and the other begin?

 

[Evo2Me]

How does one tell the difference? What if there's both a lot of grain
*and* noise - where does one stop and the other begin?

Grain is a physical entity on the film scanned by the hardware. all
instances of it will show up relatively similar - no sudden colour
change, no unexpected brightness change.

Electronic noise, OTOH, is a remnant of the scanning process itself.
It is normally present only in dark areas, and shows as unexpected
colour or brightness changes, say, a bright red pixel in the midst of
dark brown ones.

There's also the difference in shape, while pixels are all the same
size and geometric form, grain is non-uniform in these aspects.

From my experiences scanning Fuji Velvia, Kodak Ektar 125 or Kodak
Royal Gold 100 - all known to have very fine grain - with a Minolta
5400 at 5400 ppi, pixels are much smaller than grains, thus capturing
the grain rather defined. This is with and without Grain Dissolver.

 

[Kennedy McEwen]

Electronic noise, OTOH, is a remnant of the scanning process itself.
It is normally present only in dark areas, and shows as unexpected
colour or brightness changes, say, a bright red pixel in the midst of
dark brown ones.

That certainly isn't true. Noise is present throughout the brightness
range. Indeed, for a high performance scanner the noise present in the
highlights will be greater than the noise in the shadows, since a major
component of the total noise, photon noise, is proportional to the
square root of the number of photons captured by each pixel. Noise is
usually more visible and objectionable in the shadows, but that is not
because the noise is only present there, or even greater in those
regions than in the mid tones and highlights. It is more obvious in
shadows because what you are visually assessing is the signal to noise
ratio, SNR, not the noise itself.

 

[Marjolein Katsma]

Evo2Me ([email protected]) wrote in

Grain is a physical entity on the film scanned by the hardware. all
instances of it will show up relatively similar - no sudden colour
change, no unexpected brightness change.

Electronic noise, OTOH, is a remnant of the scanning process itself.
It is normally present only in dark areas, and shows as unexpected
colour or brightness changes, say, a bright red pixel in the midst of
dark brown ones.

Problem is, "theoretically" I know exactly what each one *is* (I even
spent a few years working for a photochemical company, and I've seen
electron microscope pictures of grains in film.)

But I have yet to see a scan where I can clearly see grains; I have a
few scans of very "dark" (actually light) color negatives, two stops
underexposed, and I'm expecting to see some grainyness in a 4000dpi scan
- but all I'm seeing is what looks to me like just very bad noise (which
I can explain with the fact that for the very thin negative the scanner
has to use low exposure).

In other words, I just don't know:
- what a good (not noisy) scan of grains looks like,
- or a noisy scan without grains,
- or a noisy scan with grains as well.

I haven't done many scans yet (first decent scanner just bought) but I'd
like to know what to look for.

Examples to look at somewhere might help. Any URIs?

 

[rafe b]

Examples to look at somewhere might help. Any URIs?

Try

www.terrapinphoto.com/jmdavis


Lots of unretouched film scans there,
from all sorts of film scanners.


rafe b
www.terrapinphoto.com

 

[Kennedy McEwen]

But I have yet to see a scan where I can clearly see grains; I have a
few scans of very "dark" (actually light) color negatives, two stops
underexposed, and I'm expecting to see some grainyness in a 4000dpi scan
- but all I'm seeing is what looks to me like just very bad noise (which
I can explain with the fact that for the very thin negative the scanner
has to use low exposure).

You probably won't discern grain as such on a 4000ppi scan of an
underexposed image - grain grows with density. If you have some high
speed and/or push processed B&W film then you have a reasonable chance
of seeing grain, particularly in the denser regions of the negative.

Grain, as such, doesn't actually exist in developed colour emulsion,
what you can see are dye clouds and these have a different
characteristic from silver grain.

 

[Marjolein Katsma]

rafe b (rafebATspeakeasy.net) wrote in

www.terrapinphoto.com/jmdavis


Lots of unretouched film scans there,
from all sorts of film scanners.

Interesting page (bookmarked), but the word "grain" occurs only once on
that page - and not in a description of an image. So which is the one that
shows me grain, or grain plus noise, or no grain but noise?

Still looking for a clear example to know what it grain *looks* like on a
digital scan with or without noise....

 

[Marjolein Katsma]

Kennedy McEwen ([email protected]) wrote in

You probably won't discern grain as such on a 4000ppi scan of an
underexposed image - grain grows with density. If you have some high
speed and/or push processed B&W film then you have a reasonable chance
of seeing grain, particularly in the denser regions of the negative.

OK, I'll try to find one of those and see what I get...

Thanks for the explanation.

 

[Evo2Me]

That certainly isn't true.

I should have added "for all practical purposes".

My contentions about varying brightness may need one step more for
most here. Sure, there can be noise (i.e. information not present in
the original) which differs only slightly from surrounding pixels. I
could post a number of b/w T-Max photos I scanned last week, which on
close inspection show exactly that, noise pixels slightly off (mostly
colour but also luminance).

Two things to consider:

1. Will it show in print?
2. Is it distracting?

The first consideration doesn't play into the original question but
the second does.

Kennedy, you surely see a difference [pardon the pun] between noise as
I defined it - significant [!] difference in brightness or colour from
surrounding pixels - or the one you technically correct point out?
When information in the original image come out at 0,0,0 and there's a
pixel in there 1,1,1 - does it matter, is it noise? With 12,15,28 and
160,0,250 as the values it is obvious, isn't it?

 

[Evo2Me]

But I have yet to see a scan where I can clearly see grains

I have some here, from films I never thought I see grain in [Kodak
Ektar 125, Fuji Velvia, Kodak Royal Gold 100] and from T-Max. All
scanned with a Minolta 5400 at 5400 (with and without GD).

Luckily the grain is not visible anymore on typical prints through my
Epson 2100; only if I go really overboard size-wise grain will become
clearly visible and possibly distracting.

 

[Kennedy McEwen]

Kennedy, you surely see a difference [pardon the pun] between noise as
I defined it - significant [!] difference in brightness or colour from
surrounding pixels - or the one you technically correct point out?
When information in the original image come out at 0,0,0 and there's a
pixel in there 1,1,1 - does it matter, is it noise? With 12,15,28 and
160,0,250 as the values it is obvious, isn't it?

That depends on what you intend to do with the image. Just being in the
mid tones or highlights of the original scan doesn't mean it will be at
the same levels in the final image. That doesn't just mean large level
shifts. Just scanning a simple negative, for example, means that all of
the pixels that are in the highlights of the negative (where the added
photon noise makes that actual noise worse, but the SNR better) are
mapped to the shadows of the final image and vice versa. Then there is
the issue of gamma, which compresses the highlights of the *final*
image, making the noise less obvious, whilst stretching the shadows.

Since your example is with 8-bit data I assume that this is in a gamma
compensated space, since noise would be the least of the problems in
8-bit linear space. If that is the case then the answer to your
question is that the 1 level of noise would be just as visible at all
levels - since in gamma compensated space each step approximates the
same perceived brightness difference at all levels.

 

[Marjolein Katsma]

Evo2Me ([email protected]) wrote in 4ax.com:

But I have yet to see a scan where I can clearly see grains

I have some here, from films I never thought I see grain in [Kodak
Ektar 125, Fuji Velvia, Kodak Royal Gold 100] and from T-Max.

Well, could you post them somewhere?

Still wondering what *grain* looks like in a scan (as opposed to noise)...
I don't care whether it comes out in print or not - just what it looks
like.