paradizelost/paperless-ngx
1
0
Fork 0
mirror of https://github.com/paperless-ngx/paperless-ngx.git synced 2025-07-30 18:27:45 -05:00
Code Issues Packages Projects Releases Wiki Activity

reworked most of the tesseract parser, better logging

Browse Source
This commit is contained in:
Jonas Winkler
2020-11-02 15:40:44 +01:00
parent a3c71ddad4
commit d15405ef56
8 changed files with 90 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files

147
src/paperless_tesseract/parsers.py
View File

@@ -8,9 +8,7 @@ import langdetect
import pyocr
from django.conf import settings
from PIL import Image
from pyocr.libtesseract.tesseract_raw import \
TesseractError as OtherTesseractError
from pyocr.tesseract import TesseractError
from pyocr import PyocrException
import pdftotext
from documents.parsers import DocumentParser, ParseError
@@ -28,8 +26,8 @@ class RasterisedDocumentParser(DocumentParser):
image, whether it's a PDF, or other graphical format (JPEG, TIFF, etc.)
"""
def __init__(self, path):
super().__init__(path)
def __init__(self, path, logging_group):
super().__init__(path, logging_group)
self._text = None
def get_thumbnail(self):
@@ -53,11 +51,7 @@ class RasterisedDocumentParser(DocumentParser):
except ParseError:
# if convert fails, fall back to extracting
# the first PDF page as a PNG using Ghostscript
self.log(
"warning",
"Thumbnail generation with ImageMagick failed, "
"falling back to Ghostscript."
)
self.log('warning', 'Thumbnail generation with ImageMagick failed, falling back to ghostscript. Check your /etc/ImageMagick-x/policy.xml!')
gs_out_path = os.path.join(self.tempdir, "gs_out.png")
cmd = [settings.GS_BINARY,
"-q",
@@ -100,9 +94,33 @@ class RasterisedDocumentParser(DocumentParser):
images = self._get_greyscale()
if not images:
raise ParseError("Empty document, nothing to do.")
try:
self._text = self._get_ocr(images)
sample_page_index = int(len(images) / 2)
self.log("info", "Attempting language detection on page {} of {}...".format(sample_page_index+1, len(images)))
sample_page_text = self._ocr([images[sample_page_index]], settings.OCR_LANGUAGE)[0]
guessed_language = self._guess_language(sample_page_text)
if not guessed_language or guessed_language not in ISO639:
self.log("warning", "Language detection failed.")
ocr_pages = self._complete_ocr_default_language(images, sample_page_index, sample_page_text)
elif ISO639[guessed_language] == settings.OCR_LANGUAGE:
self.log("info", "Detected language: {} (default language)".format(guessed_language))
ocr_pages = self._complete_ocr_default_language(images, sample_page_index, sample_page_text)
elif not ISO639[guessed_language] in pyocr.get_available_tools()[0].get_available_languages():
self.log("warning","Detected language {} is not available on this system.".format(guessed_language))
ocr_pages = self._complete_ocr_default_language(images, sample_page_index, sample_page_text)
else:
self.log("info","Detected language: {}".format(guessed_language))
ocr_pages = self._ocr(images, ISO639[guessed_language])
self.log("info", "OCR completed.")
self._text = strip_excess_whitespace(" ".join(ocr_pages))
return self._text
except OCRError as e:
raise ParseError(e)
@@ -111,6 +129,8 @@ class RasterisedDocumentParser(DocumentParser):
Greyscale images are easier for Tesseract to OCR
"""
self.log("info", "Converting document {} into greyscale images...".format(self.document_path))
# Convert PDF to multiple PNMs
pnm = os.path.join(self.tempdir, "convert-%04d.pnm")
run_convert(
@@ -127,91 +147,43 @@ class RasterisedDocumentParser(DocumentParser):
if f.endswith(".pnm"):
pnms.append(os.path.join(self.tempdir, f))
self.log("info", "Running unpaper on {} pages...".format(len(pnms)))
# Run unpaper in parallel on converted images
with Pool(processes=settings.OCR_THREADS) as pool:
pool.map(run_unpaper, itertools.product([settings.UNPAPER_BINARY], pnms))
# Return list of converted images, processed with unpaper
pnms = []
for f in os.listdir(self.tempdir):
if f.endswith(".unpaper.pnm"):
pnms.append(os.path.join(self.tempdir, f))
pnms = pool.map(run_unpaper, pnms)
return sorted(filter(lambda __: os.path.isfile(__), pnms))
def _guess_language(self, text):
try:
guess = langdetect.detect(text)
self.log("debug", "Language detected: {}".format(guess))
return guess
except Exception as e:
self.log("warning", "Language detection error: {}".format(e))
def _get_ocr(self, imgs):
"""
Attempts to do the best job possible OCR'ing the document based on
simple language detection trial & error.
"""
if not imgs:
raise OCRError("Empty document, nothing to do.")
self.log("info", "OCRing the document")
# Since the division gets rounded down by int, this calculation works
# for every edge-case, i.e. 1
middle = int(len(imgs) / 2)
raw_text = self._ocr([imgs[middle]], settings.OCR_LANGUAGE)
guessed_language = self._guess_language(raw_text)
if not guessed_language or guessed_language not in ISO639:
self.log("warning", "Language detection failed!")
raw_text = self._assemble_ocr_sections(imgs, middle, raw_text)
return raw_text
if ISO639[guessed_language] == settings.OCR_LANGUAGE:
raw_text = self._assemble_ocr_sections(imgs, middle, raw_text)
return raw_text
try:
return self._ocr(imgs, ISO639[guessed_language])
except pyocr.pyocr.tesseract.TesseractError:
self.log(
"warning",
"OCR for {} failed, but we're going to stick with what "
"we've got since FORGIVING_OCR is enabled.".format(
guessed_language
)
)
raw_text = self._assemble_ocr_sections(imgs, middle, raw_text)
return raw_text
return None
def _ocr(self, imgs, lang):
"""
Performs a single OCR attempt.
"""
if not imgs:
return ""
self.log("info", "Parsing for {}".format(lang))
self.log("info", "Performing OCR on {} page(s) with language {}".format(len(imgs), lang))
with Pool(processes=settings.OCR_THREADS) as pool:
r = pool.map(image_to_string, itertools.product(imgs, [lang]))
r = " ".join(r)
return r
# Strip out excess white space to allow matching to go smoother
return strip_excess_whitespace(r)
def _assemble_ocr_sections(self, imgs, middle, text):
def _complete_ocr_default_language(self, images, sample_page_index, sample_page):
"""
Given a `middle` value and the text that middle page represents, we OCR
the remainder of the document and return the whole thing.
"""
text = self._ocr(imgs[:middle], settings.OCR_LANGUAGE) + text
text += self._ocr(imgs[middle + 1:], settings.OCR_LANGUAGE)
return text
# text = self._ocr(imgs[:middle], settings.OCR_LANGUAGE) + text
# text += self._ocr(imgs[middle + 1:], settings.OCR_LANGUAGE)
images_copy = list(images)
del images_copy[sample_page_index]
if images_copy:
self.log('info', 'Continuing ocr with default language.')
ocr_pages = self._ocr(images_copy, settings.OCR_LANGUAGE)
ocr_pages.insert(sample_page_index, sample_page)
return ocr_pages
else:
return [sample_page]
def run_convert(*args):
@@ -225,13 +197,16 @@ def run_convert(*args):
raise ParseError("Convert failed at {}".format(args))
def run_unpaper(args):
unpaper, pnm = args
command_args = (unpaper, "--overwrite", "--quiet", pnm,
pnm.replace(".pnm", ".unpaper.pnm"))
def run_unpaper(pnm):
pnm_out = pnm.replace(".pnm", ".unpaper.pnm")
command_args = (settings.UNPAPER_BINARY, "--overwrite", "--quiet", pnm,
pnm_out)
if not subprocess.Popen(command_args).wait() == 0:
raise ParseError("Unpaper failed at {}".format(command_args))
return pnm_out
def strip_excess_whitespace(text):
collapsed_spaces = re.sub(r"([^\S\r\n]+)", " ", text)
@@ -245,14 +220,18 @@ def strip_excess_whitespace(text):
def image_to_string(args):
img, lang = args
ocr = pyocr.get_available_tools()[0]
with Image.open(os.path.join(settings.SCRATCH_DIR, img)) as f:
with Image.open(img) as f:
if ocr.can_detect_orientation():
try:
orientation = ocr.detect_orientation(f, lang=lang)
f = f.rotate(orientation["angle"], expand=1)
except (TesseractError, OtherTesseractError, AttributeError):
except Exception:
# Rotation not possible, ignore
pass
return ocr.image_to_string(f, lang=lang)
try:
return ocr.image_to_string(f, lang=lang)
except PyocrException as e:
raise OCRError(e)
def get_text_from_pdf(pdf_file):