Re: Multi-sampling and "2400x4800 dpi" scanners

Gordon Moat wrote: 
Good afternoon LjL,

[snip]

Perhaps you might want to look at the firmware. The code might be simpler, and you
may be able to affect a change in less attempts.

I would certainly like to look at the firmware!
But I don't think there is any (documented) way to look at it and/or modify it. I guess it's one of the things Epson wants to keep most secret of their scanners!

> [snip]
>

If you can figure out how to easily get it apart, you might discover a bit more.
However, dust could then become an issue. It would make a nice opportunity to
clean the other side of the glass.


Nah, it's too new to make such attempts...

[snip]

Well, if you can suggest a simple method for measuring real resolution,
I would be happy to try and find out. Well, I wouldn't necessarily be
*happy*, on a second thought.


Ted Harris wrote an article in the May/June 2005 issue of View Camera about
scanners. He used a USAF1951 test target and a T4110 step wedge; might be
something to try out. Really well written and presented article, though it would
have been nice to include larger sample images.

I don't want to spend money (or send money abroad, as is often the case with these things) on test targets. I would certainly like to know my scanner's true resolution, but I'm not going to spend money for that... after all, there's nothing I can do to improve it.

(Though I might consider getting a calibration target - that does have a practical use!)

I am currently experimenting with "slanted edge" and Imatest, and I'll publish some results soon, although I'm afraid they're going to be heavily off, as I haven't quite understood the procedure.

> [snip]
>

But keeping the image at a 2:1 ratio is not what I plan to do.
What I plan to do is to take it back to a 1:1 ratio by downsampling on
the y axis: this way I get a 2400x2400 dpi picture, which is less noisy
than the same picture taken directly at 2400x2400 dpi from the scanner.


So you want to capture a stacked rectangle in one dimension, then compress the
long end to display a square. Sounds like some pretty hefty algorithm to avoid
creating artefacts in the final image file. Wouldn't that actually reduce the edge
sharpness and resolution of any diagonals in the source image?

That's exactly my main concern.
On the other hand, just about everyone recommends to scan at a high resolution and then scale down instead of scanning directly at a lower resolution. And this is exactly what I'm doing, except that the scaling down is only in one direction in my case.

My main doubt was, what is the "best" or the "right" algorithm to do
the downsampling? I usually just use (bi)cubic resampling when I want
to resize something; but in this case, I've got some specific
information about the original data -- i.e. that each line is shifted
by half a pixel from the previous line (a quarter of a pixel actually,
since the CCD is staggered, but I think this shouldn't be important as
long as I want to keep the final image at 2400x2400).


I think Bart van der Wolf had a short test page of several algorithms. You might
want to search archives, or contact him about that. Few people ever suggest using
bilinear, though there are some images that work better using that.

I have found http://heim.ifi.uio.no/~gisle/photo/interpolation.html . Bart van der Wolf is involved, but I don't know if it's the page you meant.

But, you see, I'm not looking for the "best looking" algorithm in general -- what I'm looking for is the right algorithm to downsample things made with an half-pixel shift etc. etc.
It might even come out to be bilinear!

It would seem
that having an option to use more than one algorithm would be of greater benefit
that forcing just one to work. Of course, the programming would be much more hefty
to do that.

But my goal is to automate the scanning process, so looking at each image before storing it isn't really an option.
Storing the images in the original ratio to leave room for future decisions is also, well, not "not an option" but impractical, due to the file sizes involved.

But again, more than the algorithm that "looks best", I'm searching for the algorithm that is the most "correct" in the context I'm working with.
Hopefully, it will also be the one that looks best with most images!

> [snip] 
Ideally you try to do as much as you can prior to dumping the image file into
PhotoShop. Nearly all operations in PhotoShop are destructive editing. The other
issue is that getting the scan optimal reduces billing time in PhotoShop, though
that is more a commercial workflow necessity.

My idea is to have a script (which, in a basic form, is already in place) to batch-scan and do all the non-destructive (or destructive-but-the-file-would-be-too-large-otherwise) corrections.

The resulting images would be stored for archival.

Another script, or the same script, would create copies of the images for viewing, where the various destructive transformations *are* applied (USM, the finer histogram corrections, resizing to 1200x1200dpi, etc).

This second part would of course be performed by me in Photoshop instead of by the script, for pictures I care about particularly.
The script would just work as a "one hour photo" equivalent.

>>>> [snip]
>>>
Or how to actually still view it as a square image.

By downsampling. 

Resize on one axis. Still images sent to video NLEs need a conversion to allow
them to display properly, i.e. you want a picture with a round basketball to still
look like a round basketball when displayed on a video monitor or television.

Ok, but I'm not doing video...
Actually, I also intend to be able to display my pictures on TV (I've got a "WebTV" from my ISP), but that's a very minor concern.

Besides, aren't you talking about NTSC? I'm PAL, and I'm not sure but I think PAL pixels are square (we've got more scanlines than NTSC).

> [snip]
>

My feeling is that working on the actual Dmax would be more beneficial to final
image quality than playing with the resolving ability.

Well, multi-sampling does improve DMax AFAIK, and multi-sampling is what I'm trying to "simulate" (actually there's nothing to simulate, 2x multi-sampling is there in my scanner, it's just that it shifts the CCD a little after the first sampling...).

You can see from the "positive vs negative" thread that I'm also trying to work out the way exposure time control works in my scanner (which, technically, comes with "auto-exposure" only). Longer exposure times (or, possibly, superimposing a long exp scan to a short exp scan, as Don does) would also help DMax.

You can try simple things
like using drum scanner oil on the flat bed, though clean-up is another issue.
There have been some good articles in Reponses Photo (French) in the past about
this method, and sometimes in a few other publications.

I could try just for seeing what comes out of it, but I can't really do that normally... this scanner is used for scanning paper *** by other people here. I don't have it all for me to play with.

> [snip]
>
Sounds like you are working with SANE.

Correct. SANE also have the advantage (over VueScan and Windows software, but I don't have Window on that computer anyway) that it allows to easily use the scanner from the network.

We've got three computers here, plus the "server" that the scanner is connected to, and thanks to SANE and SaneTwain we can all scan from our own computers.

Maybe the driver would do it, but I still
think a look at the firmware might show you another option. Anyway, best of luck,
and enjoy your project.

Thanks! Well, the scans I'm getting right now aren't *so* bad for what I must do with them, so at worst I'll be left with decent scans if I fail, and will hopefully have learned something in the process.
Nothing to lose.


by LjL
ljlbox@xxxxxxxxxx
.