- unify `generate_data_from_folder_training` w/ `..._evaluation`
- instead of recreating array after every batch, just zero out
- cast image results to uint8 instead of uint16
- cast categorical results to float instead of int
- make more config_params keys dependent on each other
- re-order accordingly
- in main, initialise them (as kwarg), so sacred actually
allows overriding them by named config file
- `index_start`: re-introduce cfg key, pass to Keras `Model.fit`
as `initial_epoch`
- make config keys `index_start` and `dir_of_start_model` dependent
on `continue_training`
- improve description
- `utils.provide_patches`: split up loop into
* `utils.preprocess_img` (single img function)
* `utils.preprocess_imgs` (top-level loop)
- capture exceptions for all cases (not just some)
at top level and with informative logging
- avoid repeating / delegating config keys in several
places: only as kwargs to `preprocess_img()`
- read files into memory only once, then re-use
- improve readability (avoiding long lines, repeated code)
when parsing `PrintSpace` or `Border` from PAGE-XML,
- use `lxml` XPath instead of nested loops
- convert points to polygons directly
(instead of painting on canvas and retrieving contours)
- pass result bbox in slice notation
(instead of xywh)
when matching files in `dir_images` by XML path name stem,
* use `dict` instead of `list` to assign reliably
* filter out `.xml` files (so input directories can be mixed)
* show informative warnings for files which cannot be matched
- do not restrict TF version, but depend on tf-keras and
set `TF_USE_LEGACY_KERAS=1` to avoid Keras 3 behaviour
- relax Numpy version requirement up to v2
- relax Torch version requirement
- drop TF1 session management code
- drop TF1 config in favour of TF2 config code for memory growth
- training.*: also simplify and limit line length
- training.train: always train with TensorBoard callback
after selecting the optimum angle on the original
search range, narrow down around in the vicinity
with half the range (adding computational costs,
but gaining precision)
when passing the text region mask, do not apply erosion only
if there are more than 2 columns, but iff `not erosion_hurts`
(consistent with `find_num_col`'s expectations and making
it as easy to find the column gaps on 1 and 2-column pages
as on multi-column pages)
- `find_number_of_columns_in_document`: retain vertical separators
and pass to `find_num_col` for each vertical split
- `return_boxes_of_images_by_order_of_reading_new`: reconstruct
the vertical separators from the segmentation mask and the separator
bboxes; pass it on to `find_num_col` everywhere
- `return_boxes_of_images_by_order_of_reading_new`: no need to
try-catch `find_num_col` anymore
- `return_boxes_of_images_by_order_of_reading_new`: when a vertical
split has too few columns,
* do not raise but lower the threshold `multiplier` responsible for
allowing gaps as column boundaries
* do not pass the `num_col_classifier` (i.e. expected number of
resulting columns) of the entire page to the iterative
`find_num_col` for each existing column, but only the portion
of that span
when searching for gaps between text regions, consider the vertical
separator mask (if given): add the vertical sum of vertical separators
to the peak scores (making column detection more robust if still slighly
skewed or partially obscured by multi-column regions, but fg seps are
present)
- when analysing regions spanning across columns,
disregard tiny regions (smaller than half the median size)
- if a region spans across columns just by a tiny fraction,
and therefore is not good enough for a multi-col separator,
then it should also not be good enough for a multi-col box
maker
- avoid unnecessary `fillPoly` (we already have the mask)
- do not merge hseps if vseps interfere
- remove old criterion (based on total length of hseps)
- create new criterion (no x overlap and x close to each other)
- rename identifiers:
* `sum_dis` → `sum_xspan`
* `diff_max_min_uniques` → `tot_xspan`
* np.std / np.mean → `dev_xspan`
- remove rule cutting around the center of crossing seps
(which is unnecessary and creates small isolated seps
at the center, unrelated to the actual crossing points)
- create rule cutting hseps by vseps _prior_ to merging
- `do_order_of_regions`: simplify aggregating per-box orders
for paragraphs and headings to overall order passed to
`xml_reading_order`; no need for `order_and_id_of_texts`,
no need to return `id_of_texts_tot`
- `do_order_of_regions_with_model`: no need to return `region_ids`
- writer: no need to pass `id_of_texts_tot` in `build_pagexml`
(because the latter does not preserve coordinates;
it scales, even when resizing the image;
this caused coordinate problems when matching deskewed contours)
- reduce `sigma` for smoothing of input to `find_peaks`
(so we get deeper gaps between columns)
- allow column boundaries closer to the margins
(50 instead of 100 or 200 px, 170 instead of 370 px)
- allow column boundaries closer to each other
(300 instead of 400 px)
- add a secondary `grenze` criterion for depth of gap
(relative to lowest minimum, if that is smaller than
the old criterion relative to lowest maximum)
- for calls to `find_num_col` within parts of a page,
do allow unbalanced column boundaries
- rename `return_x_start_end_mothers_childs_and_type_of_reading_order`
→ `return_multicol_separators_x_start_end`, and drop all the analysis
pertaining to mother/child relationships and full-span separators,
also drop the separator unification rules;
instead of the latter, try to combine neighbouring separators more
generally: join column spans iff there is nothing in between
(which also necessitates passing the region mask), and keep only
one of every such redundant pair;
add the top (of each page part) as full-span separator up front,
and return separators already ordered by y
- `return_boxes_of_images_by_order_of_reading_new`:
- also pass regions with separators, so they do not have to be
reconstructed from the separator coordinates, and also contain
images and other non-text region types, when trying to elongate
separators to maximize their span (without introducing overlaps)
- determine connected components of the region mask, i.e. labels
and their respective bboxes, in order to
1. gain additional multi-column separators, if possible
2. avoid cutting through regions which do cross column boundaries
later on
- whenever adding a new bbox, first look up the label map to see if
there are any multi-column regions extending to the right of the
current column; if there are, then advance not just one column
to the right, but as many as necessary to avoid cutting through
these regions
- new core algorithm: iterate separators sorted by y and then column
by column, but whenever the next separator ends in the same column
as the current one or even further left, recurse (i.e. finish that
span first before continuing with the top iteration)
- `lines` → `seps` (to distinguish from textlines)
- `text_regions_p_1_n` → `text_regions_p_d` (because all other
deskewed variables are called like this)
- `pixel` → `label`
- drop connected components analysis to test overlaps between
horizontal separators and (horizontal) neighbours (introduced
in ab17a927)
- instead of converting headings to topline and baseline during
`find_number_of_columns_in_document` (introduced in 9f1595d7),
add them to the matrix unchanged, but mark as extra type
(besides horizontal and vertical separtors)
- convert headings to toplines and baselines no earlier than in
`return_boxes_of_images_by_order_of_reading_new`
- for both headings and horizontal separators, if they already
span multiple columns, check if they would overlap (horizontal)
neighbours by looking at successively larger (left and right)
intervals of columns (and pick the largest elongation which
does not introduce any overlaps)
when y slice (`top:bot`) is not a significant part of the page,
viz. less than 22% (as in `find_number_of_columns_in_document`),
avoid forcing `find_num_col` to reach `num_col_classifier`
(allows large headers not to be split up and thus better ordered)
