…ed scale

…cs/punctuation, too

instead of detecting hlines and vlines independently,
and via costly horizontal/vertical morphology operations,
analyse image by medial axis transform (skeleton and distance
transform of all connected components);
then filter components that are too compact (inner vs outer size),
also filter by statistics of distance along the skeleton: filter
if too wide on average or too variant;
then apply morphological closing to reconnect broken segments, linking
only those components that roughly extend each other in the same direction;
finally, sort by size and filter components that are too small in inner
(skeleton length) or outer size (bbox diagonal), selecting only the topmost
candidates;
propagate from skeleton to full component and then spread a little into the
background

…gradient maximum and percentile

…omponents

- when trying to partition slices by separators topologically,
  also treat pre-existing regions like separators, but
  prevent placing them into distinct partitions
- when trying to partition slices by separators morphologically,
  also merge partitions that share any significant line labels

…-shape

- discern constructor vs. processing ctxt (initLogging, setup)
- resolve start model via resmgr
- get all text regions (recursively)
- use true tempdir for extracted files
- use CWD for model output paths (instead of dist dir)
- use binarized derived image, if possible
- use ocropy nlbin (instead of OpenCV thresholding), otherwise
- skip segment if no/empty text transcription
- simplify/deduplify hierarchy levels
- improve logging and docs

- discern constructor vs. processing ctxt (initLogging)
- pass effective log level (instead of global CLI override)
- use mets_target instead of fixed `mets.xml`
- simplify

This is just a dummy commit to try to trigger circleci

…d only relative paths)

Update dockerfile

…nore extend_margins

fix: levenshtein import (follow-up to 5282092)