This is a fully neural layout analysis tool capable of extracting arbitrarily shaped lines from documents. It operates by first labelling all baselines in the document using a FCN-style deep network and then extracting a fixed environment around these baselines.

Everything is highly experimental, subject to changes without notice, and will break frequently.

Run:

::

$ pip3 install .

to install the dependencies and the command line tool. For development purposes use:

::

$ pip3 install --editable .

Training requires a directory with triplets of input images $prefix.{plain,seeds}.png. plain are RGB inputs, seeds are 8bpp baselines annotations with each non-zero value representing a single baseline.

$ blla train --validation val train