Files
OCRmyPDF/src/ocrmypdf/_optimize.py
T

457 lines
16 KiB
Python

# © 2018 James R. Barlow: github.com/jbarlow83
#
# This file is part of OCRmyPDF.
#
# OCRmyPDF is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# OCRmyPDF is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with OCRmyPDF. If not, see <http://www.gnu.org/licenses/>.
from pathlib import Path
from subprocess import CalledProcessError
import concurrent.futures
from collections import defaultdict
import struct
from io import BytesIO
from PIL import Image
from .lib import fitz
import pikepdf
from . import leptonica
from .helpers import re_symlink
from .exec import pngquant, jbig2enc
PAGE_GROUP_SIZE = 10
SIMPLE_COLORSPACES = {'/DeviceRGB', '/DeviceGray', '/CalRGB', '/CalGray'}
JPEG_QUALITY = 75
PNG_QUALITY = (65, 75)
def filter_decodeparms(obj):
"""
PDF has a lot of optional data structures concerning /Filter and
/DecodeParms. /Filter can be absent or a name or an array, /DecodeParms
can be absent or a dictionary (if /Filter is a name) or an array (if
/Filter is an array). When both are arrays the lengths match.
Normalize this into:
[(/FilterName, {/DecodeParmName: Value, ...}), ...]
If there are no filters then the return is
[('', {})]
The order of /Filter matters as indicates the encoding/decoding sequence.
"""
normalized = []
filters = []
filt = obj.get('/Filter', None)
if filt is None:
return [('', {})]
if filt.type_code == pikepdf.ObjectType.array:
filters.extend(filt)
elif filt.type_code == pikepdf.ObjectType.name:
filters.append(filt)
decodeparms = obj.get('/DecodeParms', pikepdf.Array([]))
if decodeparms.type_code == pikepdf.ObjectType.dictionary:
decodeparms = pikepdf.Array([decodeparms])
for n, dp in enumerate(decodeparms):
filt_parm = (filters[n], dp)
normalized.append(filt_parm)
if len(normalized) == 0:
for filt in filters:
filt_parm = (filt, {})
normalized.append(filt_parm)
if len(normalized) == 0:
return [('', {})]
return normalized
def generate_ccitt_header(data, w, h, decode_parms):
# https://stackoverflow.com/questions/2641770/
# https://www.itu.int/itudoc/itu-t/com16/tiff-fx/docs/tiff6.pdf
if not decode_parms:
raise ValueError("/CCITTFaxDecode without /DecodeParms")
if decode_parms.get("/K", 1) < 0:
ccitt_group = 4 # Pure two-dimensional encoding (Group 4)
else:
ccitt_group = 3
img_size = len(data)
tiff_header_struct = '<' + '2s' + 'H' + 'L' + 'H' + 'HHLL' * 8 + 'L'
tiff_header = struct.pack(
tiff_header_struct,
b'II', # Byte order indication: Little endian
42, # Version number (always 42)
8, # Offset to first IFD
8, # Number of tags in IFD
256, 4, 1, w, # ImageWidth, LONG, 1, width
257, 4, 1, h, # ImageLength, LONG, 1, length
258, 3, 1, 1, # BitsPerSample, SHORT, 1, 1
259, 3, 1, ccitt_group, # Compression, SHORT, 1, 4 = CCITT Group 4 fax encoding
262, 3, 1, 0, # Thresholding, SHORT, 1, 0 = WhiteIsZero
273, 4, 1, struct.calcsize(tiff_header_struct), # StripOffsets, LONG, 1, length of header
278, 4, 1, h, # RowsPerStrip, LONG, 1, length
279, 4, 1, img_size, # StripByteCounts, LONG, 1, size of image
0 # last IFD
)
return tiff_header
def png_name(root, xref):
return str(root / '{:08d}.png'.format(xref))
def jpg_name(root, xref):
return str(root / '{:08d}.jpg'.format(xref))
def extract_image(*, doc, pike, root, log, image, xref, jbig2s,
pngs, jpegs, options):
if image.Subtype != '/Image':
return False
if image.Length < 100:
log.debug("Skipping small image, xref {}".format(xref))
return False
bpc = int(image.get('/BitsPerComponent', 8))
cs = image.get('/ColorSpace', '')
w = int(image.Width)
h = int(image.Height)
filtdps = filter_decodeparms(image)
if len(filtdps) > 1:
log.debug("Skipping multiply filtered, xref {}".format(xref))
return False
filtdp = filtdps[0]
icc = None
indexed = False
log.debug(repr(cs))
if cs[0] == '/ICCBased':
icc = cs[1]
cs = icc.stream_dict.get('/Alternate', '')
elif cs[0] == '/Indexed':
indexed = True
cs = cs[1]
log.debug(repr(cs))
if bpc > 8:
return False # Don't mess with wide gamut images
if bpc == 1 and filtdp[0] != '/JBIG2Decode' and jbig2enc.available():
if filtdp[0] == '/CCITTFaxDecode':
data = image.read_raw_bytes()
try:
header = generate_ccitt_header(data, w, h, filtdp[1])
except ValueError as e:
log.info(e)
return False
stream = BytesIO()
stream.write(header)
stream.write(data)
stream.seek(0)
with Image.open(stream) as im:
im.save(png_name(root, xref))
else:
return False
jbig2s.append(xref)
elif filtdp[0] == '/JPXDecode':
return False
elif filtdp[0] == '/DCTDecode' \
and cs in SIMPLE_COLORSPACES \
and options.optimize >= 2:
raw_jpeg = pike._get_object_id(xref, 0)
color_transform = filtdp[1].get('/ColorTransform', 1)
if color_transform != 1:
return False # Don't mess with JPEGs other than YUV
# This is a simple heuristic derived from some training data, that has
# about a 70% chance of guessing whether the JPEG is high quality,
# and possibly recompressible, or not. The number itself doesn't mean
# anything.
# bytes_per_pixel = int(raw_jpeg.Length) / (w * h)
# jpeg_quality_estimate = 117.0 * (bytes_per_pixel ** 0.213)
# if jpeg_quality_estimate < 65:
# return False
# We could get the ICC profile here, but there's no need to look at it
# for quality transcoding
# if icc:
# stream = BytesIO(raw_jpeg.read_raw_bytes())
# iccbytes = icc.read_bytes()
# with Image.open(stream) as im:
# im.save(jpg_name(root, xref), icc_profile=iccbytes)
raw_jpeg_data = raw_jpeg.read_raw_bytes()
Path(jpg_name(root, xref)).write_bytes(raw_jpeg_data)
jpegs.append(xref)
elif indexed \
and cs in SIMPLE_COLORSPACES \
and options.optimize >= 3 \
and fitz:
# Try to improve on indexed images - these are far from low hanging
# fruit in most cases
pix = fitz.Pixmap(doc, xref)
pix.writePNG(png_name(root, xref), savealpha=False)
pngs.append(xref)
elif cs in SIMPLE_COLORSPACES and fitz:
# For any 'inferior' filter including /FlateDecode we extract
# and recode as /FlateDecode
# raw_png = pike._get_object_id(xref, 0)
# raw_png_data = raw_png.read_raw_bytes()
# (root / '{:08d}.png'.format(xref)).write_bytes(raw_png_data)
pix = fitz.Pixmap(doc, xref)
pix.writePNG(png_name(root, xref), savealpha=False)
pngs.append(xref)
else:
return False
return True
def extract_images(doc, pike, root, log, options):
# Extract images we can improve
changed_xrefs = set()
jbig2_groups = defaultdict(lambda: [])
jpegs = []
pngs = []
errors = 0
for pageno, page in enumerate(pike.pages):
group, _ = divmod(pageno, PAGE_GROUP_SIZE)
try:
xobjs = page.Resources.XObject
except AttributeError:
continue
for imname, image in dict(xobjs).items():
xref = image._objgen[0]
if xref in changed_xrefs:
continue # Don't improve same image twice
try:
result = extract_image(
doc=doc, pike=pike, root=root, log=log, image=image,
xref=xref, jbig2s=jbig2_groups[group], pngs=pngs,
jpegs=jpegs, options=options
)
if result:
changed_xrefs.add(xref)
except Exception as e:
log.debug("Image {} xref {}".format(imname, xref))
log.debug(repr(e))
errors += 1
# Elide empty groups
jbig2_groups = {group: xrefs for group, xrefs in jbig2_groups.items()
if len(xrefs) > 0}
log.debug(
"Optimizable images: "
"JBIG2 groups: {} JPEGs: {} PNGs: {} Errors: {}".format(
len(jbig2_groups), len(jpegs), len(pngs), errors
))
return changed_xrefs, jbig2_groups, jpegs, pngs
def convert_to_jbig2(pike, jbig2_groups, root, log, options):
"""
Convert a group of JBIG2 images and insert into PDF.
We use a group because JBIG2 works best with a symbol dictionary that spans
multiple pages. When inserted back into the PDF, each JBIG2 must reference
the symbol dictionary it is associated with. So convert a group at a time,
and replace their streams with a parameter set that points to the
appropriate dictionary.
If too many pages shared the same dictionary JBIG2 encoding becomes more
expensive and less efficient.
"""
with concurrent.futures.ThreadPoolExecutor(
max_workers=options.jobs) as executor:
futures = []
for group, xrefs in jbig2_groups.items():
prefix = 'group{:08d}'.format(group)
future = executor.submit(
jbig2enc.convert_group,
cwd=str(root),
infiles=(png_name(root, xref) for xref in xrefs),
out_prefix=prefix
)
futures.append(future)
for future in concurrent.futures.as_completed(futures):
proc = future.result()
log.debug(proc.stderr)
for group, xrefs in jbig2_groups.items():
prefix = 'group{:08d}'.format(group)
jbig2_globals_data = (root / (prefix + '.sym')).read_bytes()
jbig2_globals = pikepdf.Stream(pike, jbig2_globals_data)
for n, xref in enumerate(xrefs):
jbig2_im_file = root / (prefix + '.{:04d}'.format(n))
jbig2_im_data = jbig2_im_file.read_bytes()
im_obj = pike._get_object_id(xref, 0)
im_obj.write(
jbig2_im_data, pikepdf.Name('/JBIG2Decode'),
pikepdf.Dictionary({
'/JBIG2Globals': jbig2_globals
})
)
def transcode_jpegs(pike, jpegs, root, log, options):
for xref in jpegs:
in_jpg = Path(jpg_name(root, xref))
opt_jpg = in_jpg.with_suffix('.opt.jpg')
# This produces a debug warning from PIL
# DEBUG:PIL.Image:Error closing: 'NoneType' object has no attribute
# 'close'. Seems to be mostly harmless
# https://github.com/python-pillow/Pillow/issues/1144
with Image.open(str(in_jpg)) as im:
im.save(str(opt_jpg),
optimize=True,
quality=JPEG_QUALITY)
if opt_jpg.stat().st_size > in_jpg.stat().st_size:
log.debug("xref {}, jpeg, made larger - skip".format(xref))
continue
compdata = leptonica.CompressedData.open(opt_jpg)
im_obj = pike._get_object_id(xref, 0)
im_obj.write(
compdata.read(), pikepdf.Name('/DCTDecode'),
pikepdf.Null()
)
def transcode_pngs(pike, pngs, root, options):
if options.optimize >= 2:
with concurrent.futures.ThreadPoolExecutor(
max_workers=options.jobs) as executor:
for xref in pngs:
executor.submit(
pngquant.quantize,
png_name(root, xref), png_name(root, xref),
PNG_QUALITY[0], PNG_QUALITY[1])
for xref in pngs:
im_obj = pike._get_object_id(xref, 0)
# Open, transcode (!), package for PDF
pix = leptonica.Pix.open(png_name(root, xref))
compdata = pix.generate_pdf_ci_data(leptonica.lept.L_FLATE_ENCODE, 0)
# This is what we should be doing: open the compressed data without
# transcoding. However this shifts each pixel row by one for some
# reason.
#compdata = leptonica.CompressedData.open(png_name(root, xref))
if len(compdata) > int(im_obj.stream_dict.Length):
continue # If we produced a larger image, don't use
predictor = pikepdf.Null()
if compdata.predictor > 0:
predictor = pikepdf.Dictionary({'/Predictor': compdata.predictor})
im_obj.BitsPerComponent = compdata.bps
im_obj.Width = compdata.w
im_obj.Height = compdata.h
if compdata.ncolors > 0:
palette_pdf_string = compdata.get_palette_pdf_string()
palette_data = pikepdf.Object.parse(palette_pdf_string)
palette_stream = pikepdf.Stream(pike, bytes(palette_data))
palette = [pikepdf.Name('/Indexed'), pikepdf.Name('/DeviceRGB'),
compdata.ncolors - 1, palette_stream]
cs = palette
else:
if compdata.spp == 1:
cs = pikepdf.Name('/DeviceGray')
elif compdata.spp == 3:
cs = pikepdf.Name('/DeviceRGB')
elif compdata.spp == 4:
cs = pikepdf.Name('/DeviceCMYK')
im_obj.ColorSpace = cs
im_obj.write(compdata.read(), pikepdf.Name('/FlateDecode'), predictor)
def optimize(
input_file,
output_file,
log,
context):
options = context.get_options()
if options.optimize == 0:
re_symlink(input_file, output_file, log)
return
global PNG_QUALITY
global JPEG_QUALITY
if options.optimize == 3:
PNG_QUALITY = (20, 40)
JPEG_QUALITY = 40
if fitz:
doc = fitz.open(input_file)
else:
doc = None
pike = pikepdf.Pdf.open(input_file)
root = Path(output_file).parent / 'images'
root.mkdir(exist_ok=True)
changed_xrefs, jbig2_groups, jpegs, pngs = extract_images(
doc, pike, root, log, options)
convert_to_jbig2(pike, jbig2_groups, root, log, options)
transcode_jpegs(pike, jpegs, root, log, options)
transcode_pngs(pike, pngs, root, options)
# Not object_stream_mode + preserve_pdfa generates noncompliant PDFs
target_file = output_file + '_opt.pdf'
pike.save(target_file, preserve_pdfa=True)
input_size = Path(input_file).stat().st_size
output_size = Path(target_file).stat().st_size
ratio = input_size / output_size
savings = 1 - output_size / input_size
log.info("Optimize ratio: {:.2f} savings: {:.1f}%".format(
ratio, 100 * savings))
if savings < 0:
log.info("Optimize did not improve the file - discarded")
re_symlink(input_file, output_file, log)
else:
re_symlink(target_file, output_file, log)
if __name__ == '__main__':
import logging
import sys
from .pipeline import JobContext
from collections import namedtuple
Options = namedtuple('Options', 'jobs optimize')
logging.basicConfig(level=logging.DEBUG)
log = logging.getLogger()
ctx = JobContext()
options = Options(jobs=4, optimize=3)
ctx.set_options(options)
optimize(sys.argv[1], sys.argv[2], log, ctx)