def __init__(self, doc: fitz.Document, page: fitz.Page) -> None:
self.doc = doc
self._page = page
# Page's cropbox (to help identify what part of an image is actually
# being displayed), already rotated
self.bbox = page.bound()
# Cross-referenced images (ignore masks since they can't be easily downsampled)
all_xref_images = list(map(self._build_xref_image, page.get_images(full=True)))
smask_xrefs = set(filter(None, map(lambda xref_image: xref_image['smask'], all_xref_images)))
self.xref_images = list(filter(lambda xref_image: xref_image['xref'] not in smask_xrefs, all_xref_images))
# Lazy, memoized attributes
# Match block numbers to image hashes
self._block_hashes = None
def _analyze_page(self, page: fitz.Page):
"""Analyzes `page` and records the data extracted from it. Does
nothing if the page cannot be analyzed successfully.
"""
original_text = page.get_text() # type: ignore
if (
total_image_area(page) / page.bound().getArea()
< self.image_area_thresh
and not len([a for a in original_text if a == '�'])
> self.max_unreadable
):
metadata, orientation_used, scale = None, None, None
language = detected_language(original_text)
self.texts.append(original_text)
self.mean_confidences.append(None)
used_original_text = True
else:
metadata, orientation_used, language, scale = self._run_ocr(
page,
(detected_language(original_text)
if len(original_text) >= self.text_len_thresh
else self.languages.items[0])
)
if mean_conf(metadata) < self.coarse_thresh:
warnings.warn('Failed to analyze image.')
self.texts.append(data_to_string(
metadata.corrected if 'corrected' in metadata.columns
else metadata.text
))
self.mean_confidences.append(mean_conf(metadata))
used_original_text = False
self.languages.add_weight(language)
self.metadata.append(metadata)
self.orientations.append(orientation_used)
self.page_languages.append(language)
self.used_original_texts.append(used_original_text)
self.times.append(time.time())
self.scales.append(scale)