Advice on automatically stitching vertical scan rows of microfiche cards

[Merlijn B.W. Wajer]

Hi,

I'm working on preserving documents stored in microfiche cards. The
hardware scanners scans the cards in several rows at relatively high
resolution. Sometimes the content on the card is of such a nature that
the rows need to be stitched together vertically (think of a high
resolution map). (Actually the images also need 'stitching' together
horizontally to make a row, but this can be done pixel perfect and has
nothing to do with panotools/hugin)

There are some inherent challenges to this approach - the obvious one
being that the images are large -- a stitched together image can easily
exceed 60000x40000 pixels.

I started reading this tutorial, since I'm dealing with scans much like
flatbed scans: https://hugin.sourceforge.io/tutorials/scans/en.shtml

However, I am running into some problems. This is my current script [1],
which I hacked together using some trial and error (and various online
resources). Stitching works on reduced resolutions, but not always on
*full* resolutions, giving the following clear error message:

> enblend: warning: some images are redundant and will not be blended
> enblend: note: usually this means that at least one of the images
> enblend: note: does not belong to the set

So I would really appreciate some help/tips on how to deal with this,
and perhaps comments on whether my scripts [1] makes sense at all.

The images are in order (top to bottom), but I am not sure if this fact
is being used cpfind - I am using linearmatch.

The error message above makes me think not enough control points were
found, and I think ideally most of the control point finding would run
at reduced resolution -- which it apparently already does by default [2]
- perhaps it can help if cpfind would downscale the further (so perhaps
at 25% of the width/height)?

Contrary to the documentation, it seems to find less control points at
higher resolutions in my case, causing incomplete stitches.

I have uploaded some example input images here as PNG (with the
resolution greatly reduced, so these should stitch together just fine
with my script):

https://archive.org/~merlijn/microfiche/ (with file names
"stitchprefix*.png")

the original scans as JPEG2000 are also online, here:

https://archive.org/~merlijn/microfiche/IA1177302-01-0054_jp2/

but those require horizontal concatenation to get the row images.

Would appreciate any help or advice,
Regards,
Merlijn


[1] https://archive.org/~merlijn/microfiche/script.txt

[2]
> EXTENDED OPTIONS
> Feature description
> For speed reasons cpfind is using images, which are scaled to
their half width and height, to
> find keypoints. With the switch --fullscale cpfind is working
on the full scale images. This
> takes longer but can provide "better" and/or more control points.
>