def _recognize_baseline_line(self, line):
if self.tags_ignore is not None:
for tag in line['lines'][0]['tags'].values():
if tag in self.tags_ignore:
logger.info(
f'Ignoring line segment with tags {line["lines"][0]["tags"]} based on {tag}.'
)
return ocr_record('', [], [], line['lines'][0])
try:
box, coords = next(extract_polygons(self.im, line))
except KrakenInputException as e:
logger.warning(f'Extracting line failed: {e}')
return ocr_record('', [], [], line['lines'][0])
self.box = box
tag, net = self._resolve_tags_to_model(coords['tags'], self.nets)
# check if boxes are non-zero in any dimension
if 0 in box.size:
logger.warning(
f'bbox {coords} with zero dimension. Emitting empty record.')
return ocr_record('', [], [], coords)
# try conversion into tensor
try:
line = self.ts[tag](box)
except Exception:
return ocr_record('', [], [], coords)
# check if line is non-zero
if line.max() == line.min():
return ocr_record('', [], [], coords)
preds = net.predict(line.unsqueeze(0))[0]
# calculate recognized LSTM locations of characters
# scale between network output and network input
self.net_scale = line.shape[2] / net.outputs.shape[2]
# scale between network input and original line
self.in_scale = box.size[0] / (line.shape[2] - 2 * self.pad)
# XXX: fix bounding box calculation ocr_record for multi-codepoint labels.
pred = ''.join(x[0] for x in preds)
pos = []
conf = []
for _, start, end, c in preds:
pos.append(
compute_polygon_section(
coords['baseline'], coords['boundary'],
self._scale_val(start, 0, self.box.size[0]),
self._scale_val(end, 0, self.box.size[0])))
conf.append(c)
if self.bidi_reordering:
logger.debug('BiDi reordering record.')
return bidi_record(ocr_record(pred, pos, conf, coords),
base_dir=self.bidi_reordering
if self.bidi_reordering in ('L', 'R') else None)
else:
logger.debug('Emitting raw record')
return ocr_record(pred, pos, conf, coords)
def _recognize_baseline_line(self, line):
try:
box, coords = next(extract_polygons(self.im, line))
except KrakenInputException as e:
logger.warning(f'Extracting line failed: {e}')
return ocr_record('', [], [], line['lines'][0])
script = coords['script']
# check if boxes are non-zero in any dimension
if 0 in box.size:
logger.warning(
'bbox {} with zero dimension. Emitting empty record.'.format(
coords))
return ocr_record('', [], [], coords)
# try conversion into tensor
try:
line = self.ts[script](box)
except Exception:
return ocr_record('', [], [], coords)
# check if line is non-zero
if line.max() == line.min():
return ocr_record('', [], [], coords)
preds = self.nets[script].predict(line.unsqueeze(0))[0]
# calculate recognized LSTM locations of characters
# scale between network output and network input
net_scale = line.shape[2] / self.nets[script].outputs.shape[2]
# scale between network input and original line
in_scale = box.size[0] / (line.shape[2] - 2 * self.pad)
def _scale_val(val, min_val, max_val):
return int(
round(
min(
max(((val * net_scale) - self.pad) * in_scale,
min_val), max_val - 1)))
# XXX: fix bounding box calculation ocr_record for multi-codepoint labels.
pred = ''.join(x[0] for x in preds)
pos = []
conf = []
for _, start, end, c in preds:
pos.append(
compute_polygon_section(coords['baseline'], coords['boundary'],
_scale_val(start, 0, box.size[0]),
_scale_val(end, 0, box.size[0])))
conf.append(c)
if self.bidi_reordering:
logger.debug('BiDi reordering record.')
rec = bidi_record(ocr_record(pred, pos, conf, coords))
return rec
else:
logger.debug('Emitting raw record')
return ocr_record(pred, pos, conf, coords)
def forced_align(doc, model):
"""
Performs a forced character alignment of text with recognition model
output activations.
Argument:
doc (dict): Parsed document.
model (kraken.lib.model.TorchSeqRecognizer): Recognition model to use for alignment.
Returns:
A list of kraken.rpred.ocr_record.
"""
im = Image.open(doc['image'])
predictor = rpred.rpred(model, im, doc)
records = []
for line in doc['lines']:
bidi_text = get_display(line['text'])
gt_fst = fst_from_text(bidi_text, model.codec)
next(predictor)
lat_fst = fst_from_lattice(model.outputs)
composed_graph = _compose(lat_fst, gt_fst)
short_graph = _shortest_path(composed_graph)
pred = []
pos = []
conf = []
for act, label in _generate_line_record(short_graph,
model.outputs.shape[2] - 1):
pos.append(
compute_polygon_section(
line['baseline'], line['boundary'],
predictor._scale_val(act[0], 0, predictor.box.size[0]),
predictor._scale_val(act[1], 0, predictor.box.size[0])))
conf.append(1.0)
pred.append(model.codec.decode([(label, 0, 0, 0)])[0][0])
records.append(
rpred.bidi_record(rpred.ocr_record(pred, pos, conf, line)))
return records