Epson 4870 Resolution test

[Kennedy McEwen]

I sense that is more a point of semantics than disagreement with what I
wrote ;-)

As per ISO 16067-2 definition:
resolution: a measure of the ability of a digital image capture system, or a
component of a digital image capture system, to depict picture detail.

I'm not sure what you mean by "classically", but it sounds a bit like
semantics ;-)

By classically, I mean the classical definition of resolution as used by
Rayleigh et al.
"the ability of an optical instrument to render two closely spaced point
sources as distinct rather than a single extended source".

The ability of an optical instrument to reproduce picture detail does
not, in itself, demonstrate resolution. An example of an optical
instrument is your eye. An example of an image element depicted by your
eye yet clearly unresolved is each and every star you can see in the
night sky. Aliased information is, by the classical definition,
unresolved.

 

[Bart van der Wolf]

SNIP

No dispute Bart, but in the staggered HyperCCD case, your hypothetical
perfect pixel will be blurred over 3x3 pixels due to sample overlap.

That's what one would expect from a good lens. In practice, the transition
can be much larger as shown in this graph
http://www.xs4all.nl/~bvdwolf/main/downloads/EP2450-EdgeProfile.gif of a
sharp edge. I mounted a razor-blade in a slide mount at a 5 degree slant. I
scanned the edge almost perpendicular to the fast scan direction at
2400/1200/600 ppi, and scaled the lower resolutions up (nearest neighbor
interpolation) to match the physical size. I then took a sample profile of
the edge's brightness transition. In all three cases it took then about 7
sample positions (74 micron) to traverse the sharp edge, with the 2400 ppi
scan as the more accurate one. To me this indicates that the physical image
projection is that large. Denser sampling only increases the accuracy, not
the detail.

The 1200 and 600 ppi have significant modulation above Nyquist (and will
thus aliase), but that has to do with the very small sample fill factor
caused by skipping 1 (less than 25% area) or 2 (less than 11% area) pixels
between samples.

By the way, as a comparison, my DSE5400 needs on average less than 3 (5400
ppi) sample
positions (14 micron) to traverse the edge (e.g. 254, 50, 14, 9) if we
assume more than a magnitude of brightness to be enough contrast.

One can calculate an Edge Spread Function (ESF) by differentiating many of
such edge transitions. The slanted edge and resulting ESFs can then be used
to calculate MTF curves. If we assume a 10% modulation as approximating the
visual resolution limit, I get the following approximate resolutions for the
different sample densities:
2400 ppi - 33 cy/mm, 1200 ppi - 26 cy/mm, and 600 ppi - 18 cy/mm.

As can be seen, the resolution in cy/mm doesn't drop by a factor of 2. This
is due to the lower fill factor and the lack of lens resolution. It also
shows that one needs to measure actual resolution, because it is not
possible to derive it directly from the sample density alone.

As a comparison, my DSE5400 provided:
5400 ppi - 140 cy/mm, 2700 ppi - 75 cy/mm, 1350 ppi - 44 cy/mm.
Here the resolution drop is more in line with the lower sampling frequency.

SNIP

This is a sharpness issue too. So my only point is just to wonder why
you say "mainly the lens"? (one is reminded of Occam's Razor here <g>)

The above 7 pixel transition of a sharp edge indicates a less than stellar
lens performance. Also, the half sensel staggered overlap is not between
100% fill factor sensels. Even an 80% fill factor only overlaps 20% per
sensel (in a simplified square layout). A 70% fill factor only overlaps 5%.
What's more, scanning at 1200 ppi or less will almost guarantee zero overlap
on my Epson 2450, so potential overlap can only be a small factor of the
total performance.
I guess only manufacturers like Epson can answer with certainty how much
overlap there is in the various sensors.

Bart

 

[Wayne Fulton]

That's what one would expect from a good lens. In practice, the transition
can be much larger as shown in this graph
http://www.xs4all.nl/~bvdwolf/main/downloads/EP2450-EdgeProfile.gif of a
sharp edge. I mounted a razor-blade in a slide mount at a 5 degree slant. I
scanned the edge almost perpendicular to the fast scan direction at
2400/1200/600 ppi, and scaled the lower resolutions up (nearest neighbor
interpolation) to match the physical size. I then took a sample profile of
the edge's brightness transition. In all three cases it took then about 7
sample positions (74 micron) to traverse the sharp edge, with the 2400 ppi
scan as the more accurate one. To me this indicates that the physical image
projection is that large. Denser sampling only increases the accuracy, not
the detail.

I see the 7 pixels measured at the edge, but sorry, I dont follow much of the
rest. No big deal, I was just curious about the results.

The 1200 and 600 ppi have significant modulation above Nyquist (and will
thus aliase), but that has to do with the very small sample fill factor
caused by skipping 1 (less than 25% area) or 2 (less than 11% area) pixels
between samples.

Significant modulation is the measured steps on the edge?
I can see that a step result in a smooth edge would be aliasing.

....

The above 7 pixel transition of a sharp edge indicates a less than stellar
lens performance. Also, the half sensel staggered overlap is not between
100% fill factor sensels. Even an 80% fill factor only overlaps 20% per
sensel (in a simplified square layout). A 70% fill factor only overlaps 5%.
What's more, scanning at 1200 ppi or less will almost guarantee zero overlap
on my Epson 2450, so potential overlap can only be a small factor of the
total performance.
I guess only manufacturers like Epson can answer with certainty how much
overlap there is in the various sensors.

I dont get your overlap numbers. Wouldnt 70% fill be 20% overlap?

70% fill is 35% per side, with 15% gap remaining on that side.
Two 15% gaps leaves 20% overlap on that side.

| 15% | 20% | 15% | = 50%

FWIW, the HyperCCD sketch in the Japanese Epson site would suggest about 75%
fill, but it is of course only a stylized sketch, hardly decisive.

But I think you are saying that you do see a larger effect that is
independent of resolution, thus must be lens.

Some people have reported focus to be above the glass when scanning film.
I find it hard to believe focus shifts with film or resolution, but I wonder
if your report might be related to focus? Have you looked at that?

 

[Markus Plail]

Rubish!

So you can surely show us a scan where one can see the full 4800 dpi
resolution...

regards
Markus

 

[Bart van der Wolf]

SNIP

I see the 7 pixels measured at the edge, but sorry, I dont follow much
of the rest. No big deal, I was just curious about the results.

My reasoning is, if 2 samples taken at a wider interval (e.g. 1200 ppi
instead of 2400 ppi, so due to skipped samples there is no overlap)
produce a similar output as the same sample positions of a smaller
interval including some overlap, then there is little additional overlap
blurring (blur is a weighted average).

Significant modulation is the measured steps on the edge?

No, you couldn't know, but I was looking at the (MTF like) SFR graphs and
they show high modulation at frequencies higher than the Nyquist frequency.

SNIP

I dont get your overlap numbers. Wouldnt 70% fill be 20% overlap?

You are correct, I made a mistake.

100% Fill factor, 50% overlap
0000000000|0000000000|0000000000|
00000|0000000000|0000000000|00000

80% Fill factor, 30% overlap between 2 staggered pixels
|_00000000_|_00000000_|_00000000_|
0000_|_00000000_|_00000000_|_0000

60% Fill factor, 10% overlap between 2 staggered pixels
|__000000__|__000000__|__000000__|
000__|__000000__|__000000__|__000

FWIW, the HyperCCD sketch in the Japanese Epson site would
suggest about 75% fill, but it is of course only a stylized sketch,
hardly decisive.

So its potential effect is bigger than I figured, but only the likes of
Epson can confirm what it really is.

But I think you are saying that you do see a larger effect that is
independent of resolution, thus must be lens.

Correct. I don't deny there may be an overlap effect at the highest
resolution, but I can only assume (from my other tests) it is limited in
comparison to the effect of lens. The fill factor can of course differ with
each sensor, I only tested with the one I have.

Some people have reported focus to be above the glass when
scanning film. I find it hard to believe focus shifts with film or
resolution, but I wonder if your report might be related to focus?
Have you looked at that?

Yes, I checked the plane of optimal focus and it is approx. 1.5mm (half a
slide mount) above the platen. The depth of field is reasonably large, so
between model variation is expected to be of minor importance (unless a unit
is out-of-specs).

Bart

 

[Wayne Fulton]

Yes, I checked the plane of optimal focus and it is approx. 1.5mm (half a
slide mount) above the platen. The depth of field is reasonably large, so
between model variation is expected to be of minor importance (unless a unit
is out-of-specs).

OK, thanks, I was just grabbing at straws <g> The HyperCCD design doesnt
appear to change the lens requirements (field coverage), so I dont know why
it would be difficult.

I've never been convinced of any advantage of the staggered CCD design,
except that it obviously works well for marketing. I assume the current high
resolutions on a flatbed (covering 8.5 inches of width) must be beyond the
state of the art otherwise, without the staggered arrays. Half-size
conventional cells are straight forward, but noisy. A double-length chip
with conventional cell widths would require a monster lens (focal length and
diameter) to cover the double width at a good f stop, like comparing a medium
film camera lens to a 4x5 inch camera lens. I suspect some intermediate try
would have both problems <g>

There are a couple of high end flatbed models that zoom the lens to cover
half bed width at double resolution, with the same one conventional CCD chip.
Expensive, but surely the cost of these could come way down if sold in
consumer volume. This seems like marketings next good opportunity, something
that actually works. <g>