def degrade_line(im, eta=0, alpha=1.7, beta=1.7, alpha_0=1, beta_0=1):
"""
Degrades a line image by adding noise
Args:
im (PIL.Image): Input image
Returns:
PIL.Image in mode 'L'
"""
im = pil2array(im)
im = np.amax(im) - im
im = im * 1.0 / np.amax(im)
# foreground distance transform and flipping to white probability
fg_dist = distance_transform_cdt(im, metric='taxicab')
fg_prob = alpha_0 * np.exp(-alpha * (fg_dist**2)) + eta
fg_prob[im == 0] = 0
fg_flip = np.random.binomial(1, fg_prob)
# background distance transform and flipping to black probability
bg_dist = distance_transform_cdt(1 - im, metric='taxicab')
bg_prob = beta_0 * np.exp(-beta * (bg_dist**2)) + eta
bg_prob[im == 1] = 0
bg_flip = np.random.binomial(1, bg_prob)
# flip
im -= fg_flip
im += bg_flip
# use a circular kernel of size 3
sel = np.array([[1, 1], [1, 1]])
im = binary_closing(im, sel)
return array2pil(255 - im.astype('B') * 255)
def nlbin(im, threshold=0.5, zoom=0.5, escale=1.0, border=0.1, perc=80,
range=20, low=5, high=90):
"""
Performs binarization using non-linear processing.
Args:
im (PIL.Image):
threshold (float):
zoom (float): Zoom for background page estimation
escale (float): Scale for estimating a mask over the text region
border (float): Ignore this much of the border
perc (int): Percentage for filters
range (int): Range for filters
low (int): Percentile for black estimation
high (int): Percentile for white estimation
Returns:
PIL.Image containing the binarized image
"""
if im.mode == '1':
return im
raw = pil2array(im)
# rescale image to between -1 or 0 and 1
raw = raw/np.float(np.iinfo(raw.dtype).max)
if raw.ndim == 3:
raw = np.mean(raw, 2)
# perform image normalization
if np.amax(raw) == np.amin(raw):
raise KrakenInputException('Image is empty')
image = raw-np.amin(raw)
image /= np.amax(image)
m = interpolation.zoom(image, zoom)
m = filters.percentile_filter(m, perc, size=(range, 2))
m = filters.percentile_filter(m, perc, size=(2, range))
m = interpolation.zoom(m, 1.0/zoom)
w, h = np.minimum(np.array(image.shape), np.array(m.shape))
flat = np.clip(image[:w, :h]-m[:w, :h]+1, 0, 1)
# estimate low and high thresholds
d0, d1 = flat.shape
o0, o1 = int(border*d0), int(border*d1)
est = flat[o0:d0-o0, o1:d1-o1]
# by default, we use only regions that contain
# significant variance; this makes the percentile
# based low and high estimates more reliable
v = est-filters.gaussian_filter(est, escale*20.0)
v = filters.gaussian_filter(v**2, escale*20.0)**0.5
v = (v > 0.3*np.amax(v))
v = morphology.binary_dilation(v, structure=np.ones((escale*50, 1)))
v = morphology.binary_dilation(v, structure=np.ones((1, escale*50)))
est = est[v]
lo = np.percentile(est.ravel(), low)
hi = np.percentile(est.ravel(), high)
flat -= lo
flat /= (hi-lo)
flat = np.clip(flat, 0, 1)
bin = np.array(255*(flat > threshold), 'B')
return array2pil(bin)
def distort_line(im, distort=3.0, sigma=10, eps=0.03, delta=0.3):
"""
Distorts a line image.
Run BEFORE degrade_line as a white border of 5 pixels will be added.
Args:
im (PIL.Image): Input image
distort (float):
sigma (float):
eps (float):
delta (float):
Returns:
PIL.Image in mode 'L'
"""
w, h = im.size
# XXX: determine correct output shape from transformation matrices instead
# of guesstimating.
logger.debug('Pasting source image into canvas')
image = Image.new('L', (int(1.5 * w), 4 * h), 255)
image.paste(im, (int((image.size[0] - w) / 2), int(
(image.size[1] - h) / 2)))
line = pil2array(image.convert('L'))
# shear in y direction with factor eps * randn(), scaling with 1 + eps *
# randn() in x/y axis (all offset at d)
logger.debug('Performing affine transformation')
m = np.array([[1 + eps * np.random.randn(), 0.0],
[eps * np.random.randn(), 1.0 + eps * np.random.randn()]])
c = np.array([w / 2.0, h / 2])
d = c - np.dot(m, c) + np.array(
[np.random.randn() * delta,
np.random.randn() * delta])
line = affine_transform(line,
m,
offset=d,
order=1,
mode='constant',
cval=255)
hs = gaussian_filter(np.random.randn(4 * h, int(1.5 * w)), sigma)
ws = gaussian_filter(np.random.randn(4 * h, int(1.5 * w)), sigma)
hs *= distort / np.amax(hs)
ws *= distort / np.amax(ws)
def _f(p):
return (p[0] + hs[p[0], p[1]], p[1] + ws[p[0], p[1]])
logger.debug('Performing geometric transformation')
im = array2pil(geometric_transform(line, _f, order=1, mode='nearest'))
logger.debug('Cropping canvas to content box')
im = im.crop(ImageOps.invert(im).getbbox())
return im
def ocropy_degrade(im, distort=1.0, dsigma=20.0, eps=0.03, delta=0.3, degradations=[(0.5, 0.0, 0.5, 0.0)]):
"""
Degrades and distorts a line using the same noise model used by ocropus.
Args:
im (PIL.Image): Input image
distort (float):
dsigma (float):
eps (float):
delta (float):
degradations (list): list returning 4-tuples corresponding to
the degradations argument of ocropus-linegen.
Returns:
PIL.Image in mode 'L'
"""
w, h = im.size
# XXX: determine correct output shape from transformation matrices instead
# of guesstimating.
image = Image.new('L', (int(1.5*w), 4*h), 255)
image.paste(im, (int((image.size[0] - w) / 2), int((image.size[1] - h) / 2)))
a = pil2array(image.convert('L'))
(sigma,ssigma,threshold,sthreshold) = degradations[np.random.choice(len(degradations))]
sigma += (2*np.random.rand()-1)*ssigma
threshold += (2*np.random.rand()-1)*sthreshold
a = a*1.0/np.amax(a)
if sigma>0.0:
a = gaussian_filter(a,sigma)
a += np.clip(np.random.randn(*a.shape)*0.2,-0.25,0.25)
m = np.array([[1+eps*np.random.randn(),0.0],[eps*np.random.randn(),1.0+eps*np.random.randn()]])
w,h = a.shape
c = np.array([w/2.0,h/2])
d = c-np.dot(m, c)+np.array([np.random.randn()*delta, np.random.randn()*delta])
a = affine_transform(a, m, offset=d, order=1, mode='constant', cval=a[0,0])
a = np.array(a>threshold,'f')
[[r,c]] = find_objects(np.array(a==0,'i'))
r0 = r.start
r1 = r.stop
c0 = c.start
c1 = c.stop
a = a[r0-5:r1+5,c0-5:c1+5]
if distort > 0:
h,w = a.shape
hs = np.random.randn(h,w)
ws = np.random.randn(h,w)
hs = gaussian_filter(hs, dsigma)
ws = gaussian_filter(ws, dsigma)
hs *= distort/np.amax(hs)
ws *= distort/np.amax(ws)
def f(p):
return (p[0]+hs[p[0],p[1]],p[1]+ws[p[0],p[1]])
a = geometric_transform(a, f, output_shape=(h,w), order=1, mode='constant', cval=np.amax(a))
im = array2pil(a).convert('L')
return im
def degrade_line(im, eta=0.0, alpha=1.5, beta=1.5, alpha_0=1.0, beta_0=1.0):
"""
Degrades a line image by adding noise.
For parameter meanings consult [1].
Args:
im (PIL.Image): Input image
eta (float):
alpha (float):
beta (float):
alpha_0 (float):
beta_0 (float):
Returns:
PIL.Image in mode '1'
"""
logger.debug(u'Inverting and normalizing input image')
im = pil2array(im)
im = np.amax(im)-im
im = im*1.0/np.amax(im)
logger.debug(u'Calculating foreground distance transform')
fg_dist = distance_transform_cdt(1-im, metric='taxicab')
logger.debug(u'Calculating flip to white probability')
fg_prob = alpha_0 * np.exp(-alpha * (fg_dist**2)) + eta
fg_prob[im == 1] = 0
fg_flip = np.random.binomial(1, fg_prob)
logger.debug(u'Calculating background distance transform')
bg_dist = distance_transform_cdt(im, metric='taxicab')
logger.debug(u'Calculating flip to black probability')
bg_prob = beta_0 * np.exp(-beta * (bg_dist**2)) + eta
bg_prob[im == 0] = 0
bg_flip = np.random.binomial(1, bg_prob)
# flip
logger.debug(u'Flipping')
im -= bg_flip
im += fg_flip
logger.debug(u'Binary closing')
sel = np.array([[1, 1], [1, 1]])
im = binary_closing(im, sel)
logger.debug(u'Converting to image')
return array2pil(255-im.astype('B')*255)
def degrade_line(im, eta=0.0, alpha=1.5, beta=1.5, alpha_0=1.0, beta_0=1.0):
"""
Degrades a line image by adding noise.
For parameter meanings consult [1].
Args:
im (PIL.Image): Input image
eta (float):
alpha (float):
beta (float):
alpha_0 (float):
beta_0 (float):
Returns:
PIL.Image in mode '1'
"""
logger.debug('Inverting and normalizing input image')
im = pil2array(im)
im = np.amax(im) - im
im = im * 1.0 / np.amax(im)
logger.debug('Calculating foreground distance transform')
fg_dist = distance_transform_cdt(1 - im, metric='taxicab')
logger.debug('Calculating flip to white probability')
fg_prob = alpha_0 * np.exp(-alpha * (fg_dist**2)) + eta
fg_prob[im == 1] = 0
fg_flip = np.random.binomial(1, fg_prob)
logger.debug('Calculating background distance transform')
bg_dist = distance_transform_cdt(im, metric='taxicab')
logger.debug('Calculating flip to black probability')
bg_prob = beta_0 * np.exp(-beta * (bg_dist**2)) + eta
bg_prob[im == 0] = 0
bg_flip = np.random.binomial(1, bg_prob)
# flip
logger.debug('Flipping')
im -= bg_flip
im += fg_flip
logger.debug('Binary closing')
sel = np.array([[1, 1], [1, 1]])
im = binary_closing(im, sel)
logger.debug('Converting to image')
return array2pil(255 - im.astype('B') * 255)
def distort_line(im, distort=3.0, sigma=10, eps=0.03, delta=0.3):
"""
Distorts a line image.
Run BEFORE degrade_line as a white border of 5 pixels will be added.
Args:
im (PIL.Image): Input image
distort (float):
sigma (float):
eps (float):
delta (float):
Returns:
PIL.Image in mode 'L'
"""
w, h = im.size
# XXX: determine correct output shape from transformation matrices instead
# of guesstimating.
logger.debug(u'Pasting source image into canvas')
image = Image.new('L', (int(1.5*w), 4*h), 255)
image.paste(im, (int((image.size[0] - w) / 2), int((image.size[1] - h) / 2)))
line = pil2array(image.convert('L'))
# shear in y direction with factor eps * randn(), scaling with 1 + eps *
# randn() in x/y axis (all offset at d)
logger.debug(u'Performing affine transformation')
m = np.array([[1 + eps * np.random.randn(), 0.0], [eps * np.random.randn(), 1.0 + eps * np.random.randn()]])
c = np.array([w/2.0, h/2])
d = c - np.dot(m, c) + np.array([np.random.randn() * delta, np.random.randn() * delta])
line = affine_transform(line, m, offset=d, order=1, mode='constant', cval=255)
hs = gaussian_filter(np.random.randn(4*h, int(1.5*w)), sigma)
ws = gaussian_filter(np.random.randn(4*h, int(1.5*w)), sigma)
hs *= distort/np.amax(hs)
ws *= distort/np.amax(ws)
def _f(p):
return (p[0] + hs[p[0], p[1]], p[1] + ws[p[0], p[1]])
logger.debug(u'Performing geometric transformation')
im = array2pil(geometric_transform(line, _f, order=1, mode='nearest'))
logger.debug(u'Cropping canvas to content box')
im = im.crop(ImageOps.invert(im).getbbox())
return im
def dewarp(normalizer: CenterNormalizer, im: Image.Image) -> Image.Image:
"""
Dewarps an image of a line using a kraken.lib.lineest.CenterNormalizer
instance.
Args:
normalizer (kraken.lib.lineest.CenterNormalizer): A line normalizer
instance
im (PIL.Image.Image): Image to dewarp
Returns:
PIL.Image containing the dewarped image.
"""
line = pil2array(im)
temp = np.amax(line)-line
temp = temp*1.0/np.amax(temp)
normalizer.measure(temp)
line = normalizer.normalize(line, cval=np.amax(line))
return array2pil(line)
def dewarp(normalizer, im):
"""
Dewarps an image of a line using a kraken.lib.lineest.CenterNormalizer
instance.
Args:
normalizer (kraken.lib.lineest.CenterNormalizer): A line normalizer
instance
im (PIL.Image): Image to dewarp
Returns:
PIL.Image containing the dewarped image.
"""
line = pil2array(im)
temp = np.amax(line) - line
temp = temp * 1.0 / np.amax(temp)
normalizer.measure(temp)
line = normalizer.normalize(line, cval=np.amax(line))
return array2pil(line)
def degrade_line(im, mean=0.0, sigma=0.001, density=0.002):
"""
Degrades a line image by adding several kinds of noise.
Args:
im (PIL.Image): Input image
mean (float): Mean of distribution for Gaussian noise
sigma (float): Standard deviation for Gaussian noise
density (float): Noise density for Salt and Pepper noise
Returns:
PIL.Image in mode 'L'
"""
im = pil2array(im)
m = np.amax(im)
im = gaussian_filter(im.astype('f')/m, 0.5)
im += np.random.normal(mean, sigma, im.shape)
flipped = np.ceil(density/2 * im.size)
coords = [np.random.randint(0, i - 1, int(flipped)) for i in im.shape]
im[coords] = 255
coords = [np.random.randint(0, i - 1, int(flipped)) for i in im.shape]
im[coords] = 0
return array2pil(np.clip(im * m, 0, 255).astype('uint8'))
def degrade_line(im, mean=0.0, sigma=0.001, density=0.002):
"""
Degrades a line image by adding several kinds of noise.
Args:
im (PIL.Image): Input image
mean (float): Mean of distribution for Gaussian noise
sigma (float): Standard deviation for Gaussian noise
density (float): Noise density for Salt and Pepper noiase
Returns:
PIL.Image in mode 'L'
"""
im = pil2array(im)
m = np.amax(im)
im = gaussian_filter(im.astype('f') / m, 0.5)
im += np.random.normal(mean, sigma, im.shape)
flipped = np.ceil(density / 2 * im.size)
coords = [np.random.randint(0, i - 1, int(flipped)) for i in im.shape]
im[coords] = 255
coords = [np.random.randint(0, i - 1, int(flipped)) for i in im.shape]
im[coords] = 0
return array2pil(np.clip(im * m, 0, 255).astype('uint8'))
def nlbin(im: Image.Image,
threshold: float = 0.5,
zoom: float = 0.5,
escale: float = 1.0,
border: float = 0.1,
perc: int = 80,
range: int = 20,
low: int = 5,
high: int = 90) -> Image:
"""
Performs binarization using non-linear processing.
Args:
im (PIL.Image.Image):
threshold (float):
zoom (float): Zoom for background page estimation
escale (float): Scale for estimating a mask over the text region
border (float): Ignore this much of the border
perc (int): Percentage for filters
range (int): Range for filters
low (int): Percentile for black estimation
high (int): Percentile for white estimation
Returns:
PIL.Image containing the binarized image
Raises:
KrakenInputException when trying to binarize an empty image.
"""
im_str = get_im_str(im)
logger.info(f'Binarizing {im_str}')
if is_bitonal(im):
logger.info(f'Skipping binarization because {im_str} is bitonal.')
return im
# convert to grayscale first
logger.debug(f'Converting {im_str} to grayscale')
im = im.convert('L')
raw = pil2array(im)
logger.debug('Scaling and normalizing')
# rescale image to between -1 or 0 and 1
raw = raw / np.float(np.iinfo(raw.dtype).max)
# perform image normalization
if np.amax(raw) == np.amin(raw):
logger.warning(f'Trying to binarize empty image {im_str}')
raise KrakenInputException('Image is empty')
image = raw - np.amin(raw)
image /= np.amax(image)
logger.debug('Interpolation and percentile filtering')
with warnings.catch_warnings():
warnings.simplefilter('ignore', UserWarning)
m = interpolation.zoom(image, zoom)
m = filters.percentile_filter(m, perc, size=(range, 2))
m = filters.percentile_filter(m, perc, size=(2, range))
mh, mw = m.shape
oh, ow = image.shape
scale = np.diag([mh * 1.0 / oh, mw * 1.0 / ow])
m = affine_transform(m, scale, output_shape=image.shape)
w, h = np.minimum(np.array(image.shape), np.array(m.shape))
flat = np.clip(image[:w, :h] - m[:w, :h] + 1, 0, 1)
# estimate low and high thresholds
d0, d1 = flat.shape
o0, o1 = int(border * d0), int(border * d1)
est = flat[o0:d0 - o0, o1:d1 - o1]
logger.debug('Threshold estimates {}'.format(est))
# by default, we use only regions that contain
# significant variance; this makes the percentile
# based low and high estimates more reliable
logger.debug('Refine estimates')
v = est - filters.gaussian_filter(est, escale * 20.0)
v = filters.gaussian_filter(v**2, escale * 20.0)**0.5
v = (v > 0.3 * np.amax(v))
v = morphology.binary_dilation(v, structure=np.ones((int(escale * 50), 1)))
v = morphology.binary_dilation(v, structure=np.ones((1, int(escale * 50))))
est = est[v]
lo = np.percentile(est.ravel(), low)
hi = np.percentile(est.ravel(), high)
flat -= lo
flat /= (hi - lo)
flat = np.clip(flat, 0, 1)
logger.debug(f'Thresholding at {threshold}')
bin = np.array(255 * (flat > threshold), 'B')
return array2pil(bin)
def read(self, page):
"""Perfoms OCR with Kraken."""
stages = page.stages
scan = stages.get("clean", None)
if scan is None:
return None
nonLetter = self.nonLetter
model = self.ensureLoaded()
blocks = page.blocks
ocrChars = []
ocrWords = []
ocrLines = []
stages["char"] = ocrChars
stages["word"] = ocrWords
stages["line"] = ocrLines
binary = pil2array(nlbin(array2pil(scan)))
for ((stripe, block), data) in blocks.items():
(left, top, right, bottom) = data["inner"]
thisBinary = binary[top:bottom, left:right]
lines = data["bands"]["main"]["lines"]
for (ln, (up, lo)) in enumerate(lines):
lln = ln + 1
roi = thisBinary[up : lo + 1]
(b, e, roi) = removeMargins(roi, keep=16)
ocrLines.append((stripe, block, lln, left + b, top + up, left + e, top + lo))
(roiH, roiW) = roi.shape[0:2]
roi = array2pil(roi)
bounds = dict(boxes=([0, 0, roiW, roiH],), text_direction=RL)
# adapt the boxes, because they corresponds to peaks of recognition,
# not to character extends
#
# See https://github.com/mittagessen/kraken/issues/184
adaptedPreds = []
for (c, (le, to, ri, bo), conf) in chain.from_iterable(
rpred(model, roi, bounds, pad=0, bidi_reordering=True)
):
if adaptedPreds:
prevPred = adaptedPreds[-1]
prevEdge = prevPred[1][0]
else:
prevEdge = roiW
correction = int(round((prevEdge - ri) / 2))
thisRi = ri + correction
if adaptedPreds:
adaptedPreds[-1][1][0] -= correction
adaptedPreds.append([c, [le, to, thisRi, bo], conf])
if adaptedPreds:
adaptedPreds[-1][1][0] = 0
# divide into words, not only on spaces, but also on punctuation
curWord = [[], []]
inWord = True
for (c, (le, to, ri, bo), conf) in adaptedPreds:
offsetW = left + b
offsetH = top + up
pos = (le + offsetW, to + offsetH, ri + offsetW, bo + offsetH)
conf = int(round(conf * 100))
ocrChars.append((stripe, block, lln, *pos, conf, c))
spaceSeen = c == " "
changeWord = not inWord and c not in nonLetter
element = (c, pos, conf)
if spaceSeen:
curWord[1].append(element)
if spaceSeen or changeWord:
if curWord[0] or curWord[1]:
ocrWords.append((stripe, block, lln, *addWord(curWord)))
curWord = [[], []]
inWord = True
continue
if inWord:
if c in nonLetter:
inWord = False
dest = 0 if inWord else 1
curWord[dest].append(element)
if curWord[0] or curWord[1]:
ocrWords.append((stripe, block, lln, *addWord(curWord)))
page.write(stage="line,word,char")
def ocropy_degrade(im,
distort=1.0,
dsigma=20.0,
eps=0.03,
delta=0.3,
degradations=((0.5, 0.0, 0.5, 0.0), )):
"""
Degrades and distorts a line using the same noise model used by ocropus.
Args:
im (PIL.Image): Input image
distort (float):
dsigma (float):
eps (float):
delta (float):
degradations (list): list returning 4-tuples corresponding to
the degradations argument of ocropus-linegen.
Returns:
PIL.Image in mode 'L'
"""
w, h = im.size
# XXX: determine correct output shape from transformation matrices instead
# of guesstimating.
logger.debug('Pasting source image into canvas')
image = Image.new('L', (int(1.5 * w), 4 * h), 255)
image.paste(im, (int((image.size[0] - w) / 2), int(
(image.size[1] - h) / 2)))
a = pil2array(image.convert('L'))
logger.debug('Selecting degradations')
(sigma, ssigma, threshold,
sthreshold) = degradations[np.random.choice(len(degradations))]
sigma += (2 * np.random.rand() - 1) * ssigma
threshold += (2 * np.random.rand() - 1) * sthreshold
a = a * 1.0 / np.amax(a)
if sigma > 0.0:
logger.debug('Apply Gaussian filter')
a = gaussian_filter(a, sigma)
logger.debug('Adding noise')
a += np.clip(np.random.randn(*a.shape) * 0.2, -0.25, 0.25)
logger.debug('Perform affine transformation and resize')
m = np.array([[1 + eps * np.random.randn(), 0.0],
[eps * np.random.randn(), 1.0 + eps * np.random.randn()]])
w, h = a.shape
c = np.array([w / 2.0, h / 2])
d = c - np.dot(m, c) + np.array(
[np.random.randn() * delta,
np.random.randn() * delta])
a = affine_transform(a,
m,
offset=d,
order=1,
mode='constant',
cval=a[0, 0])
a = np.array(a > threshold, 'f')
[[r, c]] = find_objects(np.array(a == 0, 'i'))
r0 = r.start
r1 = r.stop
c0 = c.start
c1 = c.stop
a = a[r0 - 5:r1 + 5, c0 - 5:c1 + 5]
if distort > 0:
logger.debug('Perform geometric transformation')
h, w = a.shape
hs = np.random.randn(h, w)
ws = np.random.randn(h, w)
hs = gaussian_filter(hs, dsigma)
ws = gaussian_filter(ws, dsigma)
hs *= distort / np.amax(hs)
ws *= distort / np.amax(ws)
def _f(p):
return (p[0] + hs[p[0], p[1]], p[1] + ws[p[0], p[1]])
a = geometric_transform(a,
_f,
output_shape=(h, w),
order=1,
mode='constant',
cval=np.amax(a))
im = array2pil(a).convert('L')
return im
def nlbin(im,
threshold=0.5,
zoom=0.5,
escale=1.0,
border=0.1,
perc=80,
range=20,
low=5,
high=90):
"""
Performs binarization using non-linear processing.
Args:
im (PIL.Image):
threshold (float):
zoom (float): Zoom for background page estimation
escale (float): Scale for estimating a mask over the text region
border (float): Ignore this much of the border
perc (int): Percentage for filters
range (int): Range for filters
low (int): Percentile for black estimation
high (int): Percentile for white estimation
Returns:
PIL.Image containing the binarized image
"""
if im.mode == '1':
return im
raw = pil2array(im)
# rescale image to between -1 or 0 and 1
raw = raw / np.float(np.iinfo(raw.dtype).max)
if raw.ndim == 3:
raw = np.mean(raw, 2)
# perform image normalization
if np.amax(raw) == np.amin(raw):
raise KrakenInputException('Image is empty')
image = raw - np.amin(raw)
image /= np.amax(image)
with warnings.catch_warnings():
warnings.simplefilter('ignore', UserWarning)
m = interpolation.zoom(image, zoom)
m = filters.percentile_filter(m, perc, size=(range, 2))
m = filters.percentile_filter(m, perc, size=(2, range))
m = interpolation.zoom(m, 1.0 / zoom)
w, h = np.minimum(np.array(image.shape), np.array(m.shape))
flat = np.clip(image[:w, :h] - m[:w, :h] + 1, 0, 1)
# estimate low and high thresholds
d0, d1 = flat.shape
o0, o1 = int(border * d0), int(border * d1)
est = flat[o0:d0 - o0, o1:d1 - o1]
# by default, we use only regions that contain
# significant variance; this makes the percentile
# based low and high estimates more reliable
v = est - filters.gaussian_filter(est, escale * 20.0)
v = filters.gaussian_filter(v**2, escale * 20.0)**0.5
v = (v > 0.3 * np.amax(v))
v = morphology.binary_dilation(v, structure=np.ones((int(escale * 50), 1)))
v = morphology.binary_dilation(v, structure=np.ones((1, int(escale * 50))))
est = est[v]
lo = np.percentile(est.ravel(), low)
hi = np.percentile(est.ravel(), high)
flat -= lo
flat /= (hi - lo)
flat = np.clip(flat, 0, 1)
bin = np.array(255 * (flat > threshold), 'B')
return array2pil(bin)
def nlbin(im: Image.Image,
threshold: float = 0.5,
zoom: float = 0.5,
escale: float = 1.0,
border: float = 0.1,
perc: int = 80,
range: int = 20,
low: int = 5,
high: int = 90) -> Image:
"""
Performs binarization using non-linear processing.
Args:
im (PIL.Image.Image):
threshold (float):
zoom (float): Zoom for background page estimation
escale (float): Scale for estimating a mask over the text region
border (float): Ignore this much of the border
perc (int): Percentage for filters
range (int): Range for filters
low (int): Percentile for black estimation
high (int): Percentile for white estimation
Returns:
PIL.Image containing the binarized image
Raises:
KrakenInputException when trying to binarize an empty image.
"""
im_str = get_im_str(im)
logger.info('Binarizing {}'.format(im_str))
if is_bitonal(im):
logger.info('Skipping binarization because {} is bitonal.'.format(im_str))
return im
# convert to grayscale first
logger.debug('Converting {} to grayscale'.format(im_str))
im = im.convert('L')
raw = pil2array(im)
logger.debug('Scaling and normalizing')
# rescale image to between -1 or 0 and 1
raw = raw/np.float(np.iinfo(raw.dtype).max)
# perform image normalization
if np.amax(raw) == np.amin(raw):
logger.warning('Trying to binarize empty image {}'.format(im_str))
raise KrakenInputException('Image is empty')
image = raw-np.amin(raw)
image /= np.amax(image)
logger.debug('Interpolation and percentile filtering')
with warnings.catch_warnings():
warnings.simplefilter('ignore', UserWarning)
m = interpolation.zoom(image, zoom)
m = filters.percentile_filter(m, perc, size=(range, 2))
m = filters.percentile_filter(m, perc, size=(2, range))
mh, mw = m.shape
oh, ow = image.shape
scale = np.diag([mh * 1.0/oh, mw * 1.0/ow])
m = affine_transform(m, scale, output_shape=image.shape)
w, h = np.minimum(np.array(image.shape), np.array(m.shape))
flat = np.clip(image[:w, :h]-m[:w, :h]+1, 0, 1)
# estimate low and high thresholds
d0, d1 = flat.shape
o0, o1 = int(border*d0), int(border*d1)
est = flat[o0:d0-o0, o1:d1-o1]
logger.debug('Threshold estimates {}'.format(est))
# by default, we use only regions that contain
# significant variance; this makes the percentile
# based low and high estimates more reliable
logger.debug('Refine estimates')
v = est-filters.gaussian_filter(est, escale*20.0)
v = filters.gaussian_filter(v**2, escale*20.0)**0.5
v = (v > 0.3*np.amax(v))
v = morphology.binary_dilation(v, structure=np.ones((int(escale * 50), 1)))
v = morphology.binary_dilation(v, structure=np.ones((1, int(escale * 50))))
est = est[v]
lo = np.percentile(est.ravel(), low)
hi = np.percentile(est.ravel(), high)
flat -= lo
flat /= (hi-lo)
flat = np.clip(flat, 0, 1)
logger.debug('Thresholding at {}'.format(threshold))
bin = np.array(255*(flat > threshold), 'B')
return array2pil(bin)
