Files
OCRmyPDF/ocrmypdf/leptonica.py
T

601 lines
19 KiB
Python

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
#
# © 2013-16: jbarlow83 from Github (https://github.com/jbarlow83)
#
# Python FFI wrapper for Leptonica library
import argparse
import sys
import os
import logging
from tempfile import TemporaryFile
from ctypes.util import find_library
from .lib._leptonica import ffi
from functools import lru_cache
from enum import Enum
from .helpers import fspath
lept = ffi.dlopen(find_library('lept'))
logger = logging.getLogger(__name__)
def stderr(*objs):
"""Python 2/3 compatible print to stderr.
"""
print("leptonica.py:", *objs, file=sys.stderr)
class LeptonicaErrorTrap(object):
"""Context manager to trap errors reported by Leptonica.
Leptonica's error return codes are unreliable to the point of being
almost useless. It does, however, write errors to stderr provided that is
not disabled at its compile time. Fortunately this is done using error
macros so it is very self-consistent.
This context manager redirects stderr to a temporary file which is then
read and parsed for error messages. As a side benefit, debug messages
from Leptonica are also suppressed.
"""
def __enter__(self):
from io import UnsupportedOperation
self.tmpfile = TemporaryFile()
# Save the old stderr, and redirect stderr to temporary file
sys.stderr.flush()
try:
self.copy_of_stderr = os.dup(sys.stderr.fileno())
os.dup2(self.tmpfile.fileno(), sys.stderr.fileno(),
inheritable=False)
except UnsupportedOperation:
self.copy_of_stderr = None
return
def __exit__(self, exc_type, exc_value, traceback):
# Restore old stderr
sys.stderr.flush()
if self.copy_of_stderr is not None:
os.dup2(self.copy_of_stderr, sys.stderr.fileno())
os.close(self.copy_of_stderr)
# Get data from tmpfile (in with block to ensure it is closed)
with self.tmpfile as tmpfile:
tmpfile.seek(0) # Cursor will be at end, so move back to beginning
leptonica_output = tmpfile.read().decode(errors='replace')
assert self.tmpfile.closed
assert not sys.stderr.closed
# If there are Python errors, let them bubble up
if exc_type:
logger.warning(leptonica_output)
return False
# If there are Leptonica errors, wrap them in Python excpetions
if 'Error' in leptonica_output:
if 'image file not found' in leptonica_output:
raise FileNotFoundError()
if 'pixWrite: stream not opened' in leptonica_output:
raise LeptonicaIOError()
raise LeptonicaError(leptonica_output)
return False
class LeptonicaError(Exception):
pass
class LeptonicaIOError(LeptonicaError):
pass
class RemoveColormap(Enum):
to_binary = 0
to_grayscale = 1
to_full_color = 2
based_on_src = 3
class Pix:
"""Wrapper around leptonica's PIX object.
Leptonica uses referencing counting on PIX objects. Also, many Leptonica
functions return the original object with an increased reference count
if the operation had no effect (for example, image skew was found to be 0).
This has complications for memory management in Python. Whenever Leptonica
returns a PIX object (new or old), we wrap it in this class, which
registers it with the FFI garbage collector. pixDestroy() decrements the
reference count and only destroys when the last reference is removed.
Leptonica's reference counting is not threadsafe. This class can be used
in a threadsafe manner if a Python threading.Lock protects the data.
This class treats Pix objects as immutable. All methods return new
modified objects. This allows convenient chaining:
>>> Pix.read('filename.jpg').scale((0.5, 0.5)).deskew().show()
"""
def __init__(self, pix):
self._pix = ffi.gc(pix, Pix._pix_destroy)
def __repr__(self):
if self._pix:
s = "<leptonica.Pix image size={0}x{1} depth={2} at 0x{3:x}>"
return s.format(self._pix.w, self._pix.h, self._pix.d,
int(ffi.cast("intptr_t", self._pix)))
else:
return "<leptonica.Pix image NULL>"
def _repr_png_(self):
"""iPython display hook
returns png version of image
"""
data = ffi.new('l_uint8 **')
size = ffi.new('size_t *')
err = lept.pixWriteMemPng(data, size, self._pix, 0)
if err != 0:
raise LeptonicaIOError("pixWriteMemPng")
char_data = ffi.cast('char *', data[0])
return ffi.buffer(char_data, size[0])[:]
def __getstate__(self):
data = ffi.new('l_uint32 **')
size = ffi.new('size_t *')
err = lept.pixSerializeToMemory(self._pix, data, size)
if err != 0:
raise LeptonicaIOError("pixSerializeToMemory")
char_data = ffi.cast('char *', data[0])
# Copy from C bytes to python bytes()
data_bytes = ffi.buffer(char_data, size[0])[:]
# Can now free C bytes
lept.lept_free(char_data)
return dict(data=data_bytes)
def __setstate__(self, state):
cdata_bytes = ffi.new('char[]', state['data'])
cdata_uint32 = ffi.cast('l_uint32 *', cdata_bytes)
pix = lept.pixDeserializeFromMemory(
cdata_uint32, len(state['data']))
Pix.__init__(self, pix)
def __eq__(self, other):
return self.__getstate__() == other.__getstate__()
@property
def width(self):
return self._pix.w
@property
def height(self):
return self._pix.h
@property
def depth(self):
return self._pix.d
@property
def size(self):
return (self._pix.w, self._pix.h)
@property
def info(self):
return {'dpi': (self._pix.xres, self._pix.yres)}
@property
def mode(self):
"Return mode like PIL.Image"
if self.depth == 1:
return '1'
elif self.depth >= 16:
return 'RGB'
elif not self._pix.colormap:
return 'L'
else:
return 'P'
@classmethod
def read(cls, path):
"""Load an image file into a PIX object.
Leptonica can load TIFF, PNM (PBM, PGM, PPM), PNG, and JPEG. If
loading fails then the object will wrap a C null pointer.
"""
filename = fspath(path)
with LeptonicaErrorTrap():
return cls(lept.pixRead(os.fsencode(filename)))
def write_implied_format(
self, path, jpeg_quality=0, jpeg_progressive=0):
"""Write pix to the filename, with the extension indicating format.
jpeg_quality -- quality (iff JPEG; 1 - 100, 0 for default)
jpeg_progressive -- (iff JPEG; 0 for baseline seq., 1 for progressive)
"""
filename = fspath(path)
with LeptonicaErrorTrap():
lept.pixWriteImpliedFormat(
os.fsencode(filename),
self._pix, jpeg_quality, jpeg_progressive)
def topil(self):
"Returns a PIL.Image version of this Pix"
from PIL import Image
# Leptonica manages data in words, so it implicitly does an endian
# swap. Tell Pillow about this when it reads the data.
pix = self
if sys.byteorder == 'little':
if self.mode == 'RGB':
raw_mode = 'XBGR'
elif self.mode == 'RGBA':
raw_mode = 'ABGR'
elif self.mode == '1':
raw_mode = '1;I'
pix = Pix(lept.pixEndianByteSwapNew(pix._pix))
else:
raw_mode = self.mode
pix = Pix(lept.pixEndianByteSwapNew(pix._pix))
else:
raw_mode = self.mode # no endian swap needed
size = (pix._pix.w, pix._pix.h)
bytecount = pix._pix.wpl * 4 * pix._pix.h
buf = ffi.buffer(pix._pix.data, bytecount)
stride = pix._pix.wpl * 4
im = Image.frombytes(self.mode, size, buf, 'raw', raw_mode, stride)
return im
def show(self):
return self.topil().show()
def deskew(self, reduction_factor=0):
"""Returns the deskewed pix object.
A clone of the original is returned when the algorithm cannot find a
skew angle with sufficient confidence.
reduction_factor -- amount to downsample (0 for default) when searching
for skew angle
"""
with LeptonicaErrorTrap():
return Pix(lept.pixDeskew(self._pix, reduction_factor))
def scale(self, scale_xy):
"Returns the pix object rescaled according to the proportions given."
with LeptonicaErrorTrap():
return Pix(lept.pixScale(self._pix, scale_xy[0], scale_xy[1]))
def rotate180(self):
with LeptonicaErrorTrap():
return Pix(lept.pixRotate180(ffi.NULL, self._pix))
def rotate_orth(self, quads):
"Orthographic rotation, quads: 0-3, number of clockwise rotations"
with LeptonicaErrorTrap():
return Pix(lept.pixRotateOrth(self._pix, quads))
def find_skew(self):
"""Returns a tuple (deskew angle in degrees, confidence value).
Returns (None, None) if no angle is available.
"""
with LeptonicaErrorTrap():
angle = ffi.new('float *', 0.0)
confidence = ffi.new('float *', 0.0)
result = lept.pixFindSkew(self._pix, angle, confidence)
if result == 0:
return (angle[0], confidence[0])
else:
return (None, None)
def convert_rgb_to_luminance(self):
with LeptonicaErrorTrap():
gray_pix = lept.pixConvertRGBToLuminance(self._pix)
if gray_pix:
return Pix(gray_pix)
return None
def remove_colormap(self, removal_type):
"""Remove a palette
removal_type - RemovalColormap()
"""
with LeptonicaErrorTrap():
return Pix(lept.pixRemoveColormap(self._pix, removal_type))
def otsu_adaptive_threshold(
self, tile_size=(300, 300), kernel_size=(4, 4), scorefract=0.1):
with LeptonicaErrorTrap():
sx, sy = tile_size
smoothx, smoothy = kernel_size
p_pix = ffi.new('PIX **')
result = lept.pixOtsuAdaptiveThreshold(
self._pix,
sx, sy,
smoothx, smoothy,
scorefract,
ffi.NULL,
p_pix)
if result == 0:
return Pix(p_pix[0])
else:
return None
def otsu_threshold_on_background_norm(
self, mask=None, tile_size=(10, 15), thresh=100, mincount=50,
bgval=255, kernel_size=(2, 2), scorefract=0.1):
with LeptonicaErrorTrap():
sx, sy = tile_size
smoothx, smoothy = kernel_size
if mask is None:
mask = ffi.NULL
if isinstance(mask, Pix):
mask = mask._pix
thresh_pix = lept.pixOtsuThreshOnBackgroundNorm(
self._pix,
mask,
sx, sy,
thresh, mincount, bgval,
smoothx, smoothy,
scorefract,
ffi.NULL
)
if thresh_pix == ffi.NULL:
return None
return Pix(thresh_pix)
def crop_to_foreground(
self, threshold=128, mindist=70, erasedist=30, pagenum=0,
showmorph=0, display=0, pdfdir=ffi.NULL):
with LeptonicaErrorTrap():
cropbox = Box(lept.pixFindPageForeground(
self._pix,
threshold,
mindist,
erasedist,
pagenum,
showmorph,
display,
pdfdir))
print(repr(cropbox))
cropped_pix = lept.pixClipRectangle(
self._pix,
cropbox._box,
ffi.NULL)
return Pix(cropped_pix)
def clean_background_to_white(
self, mask=None, grayscale=None, gamma=1.0, black=0, white=255):
with LeptonicaErrorTrap():
return Pix(lept.pixCleanBackgroundToWhite(
self._pix,
mask or ffi.NULL,
grayscale or ffi.NULL,
gamma,
black,
white))
def gamma_trc(self, gamma=1.0, minval=0, maxval=255):
with LeptonicaErrorTrap():
return Pix(lept.pixGammaTRC(
ffi.NULL,
self._pix,
gamma,
minval,
maxval
))
def background_norm(
self, mask=None, grayscale=None, tile_size=(10, 15), fg_threshold=60,
min_count=40, bg_val=200, smooth_kernel=(2, 1)):
with LeptonicaErrorTrap():
return Pix(lept.pixBackgroundNorm(
self._pix,
mask or ffi.NULL,
grayscale or ffi.NULL,
tile_size[0],
tile_size[1],
fg_threshold,
min_count,
bg_val,
smooth_kernel[0],
smooth_kernel[1]
))
@staticmethod
@lru_cache(maxsize=1)
def make_pixel_sum_tab8():
return lept.makePixelSumTab8()
@staticmethod
def correlation_binary(pix1, pix2):
if get_leptonica_version() < 'leptonica-1.72':
# Older versions of Leptonica (pre-1.72) have a buggy
# implementation of pixCorrelationBinary that overflows on larger
# images. Ubuntu trusty has 1.70. Ubuntu PPA
# ppa:rebuntu16/avidemux+unofficial has "leptonlib" 1.73.
pix1_count = ffi.new('l_int32 *')
pix2_count = ffi.new('l_int32 *')
pixn_count = ffi.new('l_int32 *')
tab8 = Pix.make_pixel_sum_tab8()
lept.pixCountPixels(pix1._pix, pix1_count, tab8)
lept.pixCountPixels(pix2._pix, pix2_count, tab8)
pixn = Pix(lept.pixAnd(ffi.NULL, pix1._pix, pix2._pix))
lept.pixCountPixels(pixn._pix, pixn_count, tab8)
# Python converts these int32s to larger units as needed
# to avoid overflow. Overflow happens easily here.
correlation = (
(pixn_count[0] * pixn_count[0]) /
(pix1_count[0] * pix2_count[0])
)
return correlation
else:
correlation = ffi.new('float *', 0.0)
result = lept.pixCorrelationBinary(pix1._pix, pix2._pix,
correlation)
if result != 0:
raise LeptonicaError("Correlation failed")
return correlation[0]
@staticmethod
def _pix_destroy(pix):
p_pix = ffi.new('PIX **', pix)
lept.pixDestroy(p_pix)
# print('pix destroy ' + repr(pix))
class Box:
"""Wrapper around Leptonica's BOX objects.
See class Pix for notes about reference counting.
"""
def __init__(self, box):
self._box = ffi.gc(box, Box._box_destroy)
def __repr__(self):
if self._box:
return '<leptonica.Box x={0} y={1} w={2} h={3}>'.format(
self.x, self.y, self.w, self.h)
return '<leptonica.Box NULL>'
@property
def x(self):
return self._box.x
@property
def y(self):
return self._box.y
@property
def w(self):
return self._box.w
@property
def h(self):
return self._box.h
@staticmethod
def _box_destroy(box):
p_box = ffi.new('BOX **', box)
lept.boxDestroy(p_box)
@lru_cache(maxsize=1)
def get_leptonica_version():
"""Get Leptonica version string.
Caveat: Leptonica expects the caller to free this memory. We don't,
since that would involve binding to libc to access libc.free(),
a pointless effort to reclaim 100 bytes of memory.
"""
return ffi.string(lept.getLeptonicaVersion()).decode()
def deskew(infile, outfile, dpi):
try:
pix_source = Pix.read(infile)
except LeptonicaIOError:
raise LeptonicaIOError("Failed to open file: %s" % infile)
if dpi < 150:
reduction_factor = 1 # Don't downsample too much if DPI is already low
else:
reduction_factor = 0 # Use default
pix_deskewed = pix_source.deskew(reduction_factor)
try:
pix_deskewed.write_implied_format(outfile)
except LeptonicaIOError:
raise LeptonicaIOError("Failed to open destination file: %s" % outfile)
def remove_background(infile, outfile, tile_size=(40, 60), gamma=1.0,
black_threshold=70, white_threshold=190):
try:
pix = Pix.read(infile)
except LeptonicaIOError:
raise LeptonicaIOError("Failed to open file: %s" % infile)
pix = pix.background_norm(tile_size=tile_size).gamma_trc(
gamma, black_threshold, white_threshold)
try:
pix.write_implied_format(outfile)
except LeptonicaIOError:
raise LeptonicaIOError("Failed to open destination file: %s" % outfile)
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description="Python wrapper to access Leptonica")
subparsers = parser.add_subparsers(title='commands',
description='supported operations')
parser_deskew = subparsers.add_parser('deskew')
parser_deskew.add_argument('-r', '--dpi', dest='dpi', action='store',
type=int, default=300, help='input resolution')
parser_deskew.add_argument('infile', help='image to deskew')
parser_deskew.add_argument('outfile', help='deskewed output image')
parser_deskew.set_defaults(func=deskew)
args = parser.parse_args()
if get_leptonica_version() != u'leptonica-1.69':
print("Unexpected leptonica version: %s" % getLeptonicaVersion())
args.func(args)
def test_skew_angle():
from PIL import Image, ImageDraw
from tempfile import NamedTemporaryFile
im = Image.new(mode='1', size=(1000, 1000), color=1)
draw = ImageDraw.Draw(im)
for n in range(20):
draw.line([(50, 25 + 50*n), (950, 25 + 50*n)], width=1)
del draw
test_angles = [0.1 * ang for ang in range(1, 10)] + \
[float(ang) for ang in range(1, 7)]
test_angles += [-ang for ang in test_angles]
test_angles = sorted(test_angles)
for rotate_angle in test_angles:
rotated_im = im.rotate(rotate_angle)
with NamedTemporaryFile(prefix='lept-skew', suffix='.png', delete=True) as tmpfile:
rotated_im.save(tmpfile)
pix = pixRead(tmpfile.name)
angle, confidence = pixFindSkew(pix)
print('{0} {1} {2}'.format(rotate_angle, angle, confidence), file=sys.stderr)