I've tried to make some "slanted edge" resolution tests on my Epson
RX500 scanner.

I've used the *blade* of a cutter, a cropped scan of which can be seen at
http://ljl.150m.com/scans/blade1.gif

As you can see from the picture, the edge didn't seem very regular at
all. I wonder whether the right side to use is the blade, i.e. the one
that cuts, or the other?
I might seem thick, but I don't have much confidence with real-life
English words such as "edge", "blade", "border", etc.

Anyway, I rubbed the blade a little against another metallic object, and
the resulting scan was better, as you can see at
http://ljl.150m.com/scans/blade2.gif

Both the scanner glass and the blade were cleaned before scanning.

I then cropped both scans to a part that showed the least
irregularities, and loaded it into Imatest.

In Imatest, I set the "pixels per inch" figure (in the "Plot" group of
settings) to 2400, as the scans were made at 2400 dpi.

In the "Picture" group, I exchanged "width" for "height", as Imatest
assumed that my scans would have been wider than they were high, for
some reason.
Perhaps I should have placed the blade *horizontally* (besides the
slant) instead of vertically, to test for horizontal resolution? I would
have thought not, intuitively.

Also, Imatest says in the "Picture" group
"Enter values manually if the input image has been cropped"

I ignored this advice, since I couldn't understand it. The values given
were the correct width and height of my cropped scan, besides being swapped.


You can see the results of an Imatest run on the first scan at
http://ljl.150m.com/scans/fig-blade1.gif

and those for the second scan at
http://ljl.150m.com/scans/fig-blade2.gif


Concentrating on the better scan (the "blade2" one), I also made some
experiments with unsharp masking in Photoshop before feeding the image
to Imatest.

The results for a radius=1, amount=100 application are at
http://ljl.150m.com/scans/fig-blade2USM100-1.gif

The results for a radius=4, amount=40 application are at
http://ljl.150m.com/scans/fig-blade2USM40-4.gif


Imatest can be found at http://www.imatest.com .


by LjL
ljlbox@tiscali.it

"Lorenzo J. Lucchini" <ljlbox@tiscali.it> wrote in message
news:9sDWe.31794$nT3.16717@tornado.fastwebnet.it.. .
SNIP
The 'cutting edge' is what you need, and it should be clean.
[...]

Yes, that's typical for what you want as Imatest input to analyze.
Accuracy may be improved by saving as TIFF or PNG, as Imatest can
analyze per color channel, and a GIF has a limited accuracy.
[...]

Yes, that allows to calculate absolute cycles/mm frequencies, assuming
the scanner is really taking 2400 samples/inch.

Feeding a crop to Imatest, makes it impossible to calculate a correct
'line widths per picture height' (LW/PH) metric unless the Full Scan
Area is scanned. If a crop is fed, and the maximum scan dimensions are
manually input, then the scan's LW/PH can be directly compared with
other devices for equal size output (the magnification factor from
original to output will be factored out of the equation). To reduce
manual input for full size images (like from digicams) the image is
assumed to be wider than it is high for a LW/PH metric. One can feed
full sized images and select a/several Region of Interest(s) from
that.

You test the resolution orthogonal to (i.e. across) the edge, so it
depends on the direction of carriage movement whether you test sensor
or stepper motor resolution.

See above, it has to do with the LW/PH metric so it needs to know the
maximum height of the image.
[...]

One observation, you apparently assumed/entered/kept image gamma set
to 1.0, although the image looks like a gamma adjusted one. This value
will influence the results (just like sharpening would).

If the correct gamma adjustment is selected, the Imatest CSV output in
(multiple dimensions) allows to reconstruct the Point-Spread-Function
(PSF) of the image chain that produced the image (in this case only
the scanner (lens + sensor)). That PSF can be used to 'restore'
(within the limitations of the system) the original signal, you could
say determine optimal sharpening. The optimal* sharpening can also be
determined empirically (by analyzing multiple amounts of sharpening).

* Optimal being the maximum amount of capture sharpening, without
causing edge halo. Optimal sharpening for output depends on the output
device, because it may need 'over-sharpening' to pre-compensate for
output losses'.

And, as far as I am concerned, it's recommended for technically
inclined users whom want to optimize their image/output quality, or
better understand how different imaging components influence the final
image quality, and Imatest covers more than only resolution
quantification.

Bart

Bart van der Wolf wrote:
Worry not, the scans were made as PNM and then converted to TIFF.
I only used GIF for uploading them to the website.

Well, it ought to be, at least in the dimension of the CCD...

Well, then I don't think I care too much, unless this also influences
the other results.
Does it? It's not quite clear to me from reading the Imatest help pages.

So I suppose I did it right, as I placed the blade orthogonal (though
with some angle, as instructed) to the CCD sensor. This ought to measure
the sensor resolution.

Hm, no, I didn't. I just double-checked my shell history, and I'm
positive that I've instructed the scanner to use gamma=1.0 and disable
color correction. Unless it's doing something weird that I didn't realize...

But wait, I think it's Photoshop that gamma corrects before saving
things as GIF. Well, it's not asking me anything about it, but it does
definitely appear to be doing something to the levels, by looking at the
images.

I'll fix this for the images I've uploaded. But meanwhile, the results
I've got should be correct in this respect, since they were taken from
the original TIFF images.

Now that's interesting. I was just hoping this could have been the case.
Unfortunately, I don't think I can now get Imatest to produce CSV
output, as I have finished up my trial runs, and without a figures
window already open it seems it doesn't let me open a figures file.

By the way, I think PSF has something to do with convolutions?
I've just been toying with automatic deconvolution programs - the first
one I came across was Image Analyzer, found from one of your posts, but
then I also found some more specific and even more automated ones.

If I'm correct, then the "optimal sharping" wouldn't simply be a correct
amount/radius of unsharp masking, but more like an appropriate
deconvolution matrix?

I don't care, I am only concentrating on the scanner itself for the time
being.
But I have some doubts about the empirical approach: assuming unsharp
masking is being used, one has to determine both the amount and the
radius. How can this be done empirically?

I think I can avoid edge halo (i.e. a "mountain" and a "valley" on the
top Imatest graph?) by using multiple combinations of USM amount and
radius. How do I find out which is the correct one?

Also, the "optimal sharpening" Imatest comes up with has, as you can
see, a nice high mountain and valley! I'm a bit puzzled as to why this
is so.

English is not my native language, are you subtly telling me to give up? :-)
I was curious about my scanner's resolution (I know Imatest also gives
other parameters, but one thing at a time is more than enough for me),
and thought it could be interesting for others to see the figures I
ended up with.

Also I'd like to know what you think of my scanner's resolution (keeping
in mind what it is, i.e. a 2400dpi staggered CCD flatbed scanner).
Looking at the graph, it seems that there is zero response to
frequencies well below Nyquist? Like, 4/5 of Nyquist or so, and very
marginal response starting at 3/4 or so.

(Well I shouldn't be too surprised anyway, I knew it wasn't going to
really make 2400 dpi)

OTOH, there is Imatest output for an Epson 2450 here
http://www.xs4all.nl/~bvdwolf/temp/EP2450_SFR.png
and it looks much like what I've got, mine possibly looking a little better.

The guy measured MFT10 instead of MFT30 though, should I have done so as
well? Your other post reads like I should have.


by LjL
ljlbox@tiscali.it

LONG REPLY!

"Lorenzo J. Lucchini" <ljlbox@tiscali.it> wrote in message
news:h5VWe.32762$nT3.11568@tornado.fastwebnet.it.. .
Yes, in the so-called fast-scan direction it should be very close to
what a reputable manufacturer claims. In the slow-scan direction it
may be less accurate and can change over time.

SNIP
It is not important for other metrics, those are labeled based on the
sampling density input, e.g. 2400 samples/inch.

SNIP
That's correct, an approximate 5 to 6 degree slant is best. It
provides a full phase rotation about every 10 pixels along the edge
for a single sensor element. Many samples will help to statistically
smoothen the results of grainy images around the real values, which is
beneficial for film scans.

In that case it should be okay, you've done it as it's done best.
Gamma 1.0 will eliminate inaccuracies due to incorrect gamma
assumptions. Imatest also allows to check for true gamma if you scan a
stepwedge. That way one can check if the scanner driver is indeed
giving linear gamma results when asked to do that. Also, it can verify
if the sensor is linear in the entire range (CCDs may be less than
linear as they reach saturation).

I don't think Photoshop messes with the gamma if you tell it not to
color manage it.

SNIP
No, the CSV data output is produced when selected and saved when
producing the figures after the calculation.

Yes, the PSF is the description of the blur function that spreads the
image of a point into a larger structure. Knowing that PSF (or
calculating it from the 10-90% rise Edge function data) offers the
possibility to use it as input for a deconvolution program/function.

Yes there are many programs, some of them free as you've noticed, but
the automatic ones don't always 'guess' the PSF very well. And then
there are other issues to consider, like how to avoid noise building
up with restoration.

Well, although a simple Gaussian type of sharpening can be useful for
some images, it certainly is not optimal. Actual system Blur is
usually not a simple Gaussian convolution.

One can evaluate a large number of postprocessed attempts done with
different sharpening settings applied to the original edge scan. Once
you' ve done that for your scanner then the scanner's blur
compensation s known. Imatest makes that a lot simpler, but it still
requires some work.

Trial and error, and some more trials.

By analyzing Imatest's 10-90% Edge profile data (based on the GIF), it
seems your scan has incurred a main blur that's close to a Gaussian
blur with a radius 1.55 (it is actually a more complex blur).

That "Standardized sharpening" metric is only useful for evening out
the playing field between P/S digicams, some of which apply an unknown
amount of sharpening in camera. To still allow evaluation and
comparison of those differently sharpened images, they are
theoretically sharpened to the same amount and comparison should be
more meaningful. For scanners or unsharpened images from digicams, it
can be ignored.

No, not at all. I'm just warning other readers that testing without
proper consideration for detail will give questionable results (GIGO,
garbage-in-garbage-out, still applies). Not all readers like to test
and better understand their gear, some just want to scan on automatic
and not be bothered with details. Imatest is not intended for those
scanner users.

Yes, it usually starts with curiosity ;-)

Yes, that is partly due to the type of sensor Epson uses, and it's
similar for their other models as well. At least there is no aliasing
risk.

But you now do know that it's capable of resolving almost 30 cycles/mm
(1524 ppi) at 10% modulation (approx. limiting visual resolution), and
there is nothing being resolved above 40 cy/mm (2032 ppi), so aliasing
is effectively suppressed.

The guy is me ;-) Yes, a similar result (although yours is better),
also a 2400 ppi scanner from Epson, but I probably need to clean the
bottom side of the glass platen because I used to have higher
readings.
I follow the ISO considerations for a good correlation between the
resolution at 10% modulation (or at Nyquist if that's lower) and
limiting visual resolution. Others use different values, that's why it
is user adjustable.

Bart