Epson 4870

"Bart van der Wolf" wrote:

All we know is that the Nyquist limit defines the threshold beyond
which aliasing will interfere with the result, and contrast can
approach zero at the threshold, depending on phase and prefiltering.

We know a bit more than that. The Nyquist frequency only applies to
infinitely periodic sinusoidal signals, so finite square wave patterns
just slightly below the Nyquist can't be rendered. Not even close.

"Nyquist" or N/2 differs with sampling density. Failure to include the
higher frequency components actually creates the Gibbs phenomenon, an
overshoot that enhances edge contrast...

AKA "ringing".

But we are not adding periodic sinusoidal signals of an infinitely
increasing frequency. We are sampling.

Actually, we're not. We're talking about reasonable expectations for how
many pixels it takes to represent a line pair in a digital image created by
downsampling a grossly soft scan.

Scanner CCDs also depend on fill-factor (or spot size for
drumscanners), because that will determine how high the
feature contrast will be.

In sampling theory, where one is trying to capture infinitely periodic
signals slightly below the Nyquist frequency, one must use point
samples.

That is not how a CCD array scans the image, each sensel samples an area.

Exactly. Scanning doesn't come anywhere near meeting the conditions required
for the Nyquist theorem to apply. And digital images themselves don't
either, since patterns are usually a single line or just a few cycles.

A
sensor element has a finite size, e.g. 4x4 micron, a point sample has a
*much* smaller area. An area sample can quantize to different amplitudes
depending on the percentage of area exposed.

And thus can't accurately render frequencies near the Nyquist frequency.

So basically, the theory in the vicinity of the Nyquist theory isn't
applicable. (Representing finite patterns involves much higher
frequencies.)

Sorry, that isn't true, the theory is about sampling, and since we are
sampling, we cannot ever recreate the original continuous input signal.

Huh? The above is quite true. Nyquist claims you can, under certain
conditions, accurately record frequencies up to but not including the
Nyquist frequency. Actual sampling in scanners doesn't meet the criterion of
the Nyquist theorem.

The sampling is what causes the loss, not whether it is area or point
sampled.

No, Nyquist says you can accurately reconstruct frequencies up to but not
including the Nyquist as long as you point sample.

However, that does influence the amplitude of the frequency
component, or the shape of the MTF curve.

And where it goes to zerog.

So the idea that two pixels can represent a line pair is unreasonable.

Well that obviously is not true,

Sorry: "So the idea that two pixels can _correctly_ represent an arbitrarily
positioned line pair is unreasonable."

Happier?

because there is a number of different ways
(contrasts) a line pair can be represented, but you probably mean they
can't
be imaged by a sampling system with *absolute* accuracy. I agree, but
never
said otherwise.

The issue with the 4870 (and other staggered array sensor scanners) is
that
the effective fill factor is very high, and that produces smooth edge
transitions. This is further contrast reduced by the lens, uncoated glass
platen, and internal reflections, and we also have to presume correct
positioning in a fixed focus plane.

I don't care what the issues a given the actual performance of the beast,
I want to know what's a good target resolution to downsample to.

David J. Littleboy

Tokyo, Japan

"David J. Littleboy" wrote in message
...
SNIP
We're talking about reasonable expectations for how many
pixels it takes to represent a line pair in a digital image created
by downsampling a grossly soft scan.

That would depend on:
1. The amount of sharpening before scaling down, if any, and
2. the resampling method and amount used for scaling down, and
3. the sharpening needed to compensate for resampling losses.

A scan of a theoretically perfect image can be mimicked by scanning a 5
degree slanted sharp edge, somewhat similar to the principle used by the ISO
to determine scanner resolution. A home made solution is by mounting a razor
blade in a 35mm slide mount. Another way is by mounting a piece of
aluminum/aluminium foil, folded once to form a straight edge and flattened
with careful pressure. The slant will reduce the differences between more or
less perfect alignment with the sensor array.

That will produce a target that is not yet blurred by a camera lens, so only
the scanner can introduce deterioration. After scanning you'll also notice
the amount of lens flare and internal reflection caused by the scanner
(surround the slide mount with a full glass platen mask for best results on
a flatbed scanner). Try and avoid blooming due to overexposure of the
unobstructed view of the lightsource. Then you can draw several across edge
luminance profiles (try two 90 degrees rotated scans to have both a sensor
and a stepper motor resolution) and count the number of pixels it takes e.g.
between 90% of maximum Luminance and 10% above minimum Luminance close to
the edge on the Raw gamma 1.0 data.

Try several sharpening radii with a modest amount to see which helps to
improve the steepnes of the transition, and use that for the above step 1.
Then use a resampling method that makes sense for the intended image
content, using an amount to reduce the number of transition pixels to
between two and three. Finally, sharpen to compensate for resampling losses.
That should give the best quality, and largest possible uncompromised scan
size your scanner can produce.

Applying the same steps to a film scan should also produce good quality, but
one may want to deviate a bit based on film characteristics and intended
use. If e.g. one finally needs to enlarge again for output, then why reduce
size first, and if the final image needs to be reduced further then it is
better not to resample twice. Final sharpening should be based on final
output size.

Testing the performance as described above will give a good feeling what is
needed to approach the best possible 10%-90% Luminance difference transition
width in a real image. Lower image contrasts will only lose visible contrast
faster, but the width of the transition is never larger.

As an example, my Epson 2450 @ 2400 ppi takes 10 pixels to cross the edge,
and after USM Amount 30, Radius 3.0, Threshold 0, that is reduced to 5.
Reducing that Image size to 50% gives me a 2-3 pixel transition which can be
further small radius sharpened to exaggerate the edge transition for a
visually pleasing result. A bit more sharpening in the beginning is also
possible.
In fact these settings increased the resolution from 31.5 cy/mm to 50.1
cy/mm on a sample of Fuji RA film I had scanned long ago, allowing a pretty
decent 5-6x enlargement of the film.

Bart

Thanks to everyone who replied.





"RSD99" wrote in message
...
Anybody have any real-world experience with the new Epson 4870 Scanner and either medium
format or large format negatives and/or transparencies?

"David J. Littleboy" wrote:
....
*: My opinion is that the offset CCD that Epson uses is complete garbage and
a complete waste of time, and that these scanners barely resolve 1/2 their
advertised resolution. Others are of the opion that it's a brilliant,
mathematically justified concept perfectly capable of the full advertised
resolution. Empirical results support my interpretationg.

David J. Littleboy
Tokyo, Japan

I'm with you on this one - trying scans at different resolutions I
didn't
really see any improvement going from 1600 to 3200 dpi. But for 6x6
medium
format on up to A3 paper, 1600's about enough to give a 300 dpi print,
so
fine for me. Anything larger I'd have to get printed the old fashioned
way
anyway!

Simon.