def test_dewarp_fail_color(self):
"""
Test dewarping of a color line fails
"""
with raises(ValueError):
with Image.open(resources / '000236.png') as im:
lineest.dewarp(self.lnorm, im.convert('RGB'))
def test_dewarp_gray(self):
"""
Test dewarping of a single line in grayscale
"""
with Image.open(resources / '000236.png') as im:
o = lineest.dewarp(self.lnorm, im.convert('L'))
self.assertEqual(self.lnorm.target_height, o.size[1])
def generate_input_transforms(batch: int, height: int, width: int, channels: int, pad: int) -> transforms.Compose:
"""
Generates a torchvision transformation converting a PIL.Image into a
tensor usable in a network forward pass.
Args:
batch (int): mini-batch size
height (int): height of input image in pixels
width (int): width of input image in pixels
channels (int): color channels of input
pad (int): Amount of padding on horizontal ends of image
Returns:
A torchvision transformation composition converting the input image to
the appropriate tensor.
"""
scale = 0 # type: Union[Tuple[int, int], int]
if height == 1 and width == 0 and channels > 3:
perm = (1, 0, 2)
scale = channels
mode = 'L'
# arbitrary (or fixed) height and width and channels 1 or 3 => needs a
# summarizing network (or a not yet implemented scale operation) to move
# height to the channel dimension.
elif height > 1 and width == 0 and channels in (1, 3):
perm = (0, 1, 2)
scale = height
mode = 'RGB' if channels == 3 else 'L'
# fixed height and width image => bicubic scaling of the input image, disable padding
elif height > 0 and width > 0 and channels in (1, 3):
perm = (0, 1, 2)
pad = 0
scale = (height, width)
mode = 'RGB' if channels == 3 else 'L'
elif height == 0 and width == 0 and channels in (1, 3):
perm = (0, 1, 2)
pad = 0
scale = 0
mode = 'RGB' if channels == 3 else 'L'
else:
raise KrakenInputException('Invalid input spec (variable height and fixed width not supported)')
out_transforms = []
out_transforms.append(transforms.Lambda(lambda x: x.convert(mode)))
if scale:
if isinstance(scale, int):
if mode not in ['1', 'L']:
raise KrakenInputException('Invalid mode {} for line dewarping'.format(mode))
lnorm = CenterNormalizer(scale)
out_transforms.append(transforms.Lambda(lambda x: dewarp(lnorm, x)))
out_transforms.append(transforms.Lambda(lambda x: x.convert(mode)))
elif isinstance(scale, tuple):
out_transforms.append(transforms.Resize(scale, Image.LANCZOS))
if pad:
out_transforms.append(transforms.Pad((pad, 0), fill=255))
out_transforms.append(transforms.ToTensor())
# invert
out_transforms.append(transforms.Lambda(lambda x: x.max() - x))
out_transforms.append(transforms.Lambda(lambda x: x.permute(*perm)))
return transforms.Compose(out_transforms)
def test_dewarp_bw_undewarpable(self):
"""
Test dewarping of an undewarpable line.
"""
with Image.open(
resources /
'ONB_ibn_19110701_010.tif_line_1548924556947_449.png') as im:
o = lineest.dewarp(self.lnorm, im)
self.assertEqual(self.lnorm.target_height, o.size[1])
def generate_input_transforms(batch: int,
height: int,
width: int,
channels: int,
pad: int,
valid_norm: bool = True,
force_binarization=False) -> transforms.Compose:
"""
Generates a torchvision transformation converting a PIL.Image into a
tensor usable in a network forward pass.
Args:
batch (int): mini-batch size
height (int): height of input image in pixels
width (int): width of input image in pixels
channels (int): color channels of input
pad (int): Amount of padding on horizontal ends of image
valid_norm (bool): Enables/disables baseline normalization as a valid
preprocessing step. If disabled we will fall back to
standard scaling.
force_binarization (bool): Forces binarization of input images using
the nlbin algorithm.
Returns:
A torchvision transformation composition converting the input image to
the appropriate tensor.
"""
scale = (height, width) # type: Tuple[int, int]
center_norm = False
mode = 'RGB' if channels == 3 else 'L'
if height == 1 and width == 0 and channels > 3:
perm = (1, 0, 2)
scale = (channels, 0)
if valid_norm:
center_norm = True
mode = 'L'
elif height > 1 and width == 0 and channels in (1, 3):
perm = (0, 1, 2)
if valid_norm and channels == 1:
center_norm = True
elif height == 0 and width > 1 and channels in (1, 3):
perm = (0, 1, 2)
# fixed height and width image => bicubic scaling of the input image, disable padding
elif height > 0 and width > 0 and channels in (1, 3):
perm = (0, 1, 2)
pad = 0
elif height == 0 and width == 0 and channels in (1, 3):
perm = (0, 1, 2)
pad = 0
else:
raise KrakenInputException(
'Invalid input spec {}, {}, {}, {}, {}'.format(
batch, height, width, channels, pad))
if mode != 'L' and force_binarization:
raise KrakenInputException(
'Invalid input spec {}, {}, {}, {} in'
' combination with forced binarization.'.format(
batch, height, width, channels, pad))
out_transforms = []
out_transforms.append(transforms.Lambda(lambda x: x.convert(mode)))
if force_binarization:
out_transforms.append(transforms.Lambda(lambda x: nlbin(im)))
# dummy transforms to ensure we can determine color mode of input material
# from first two transforms. It's stupid but it works.
out_transforms.append(transforms.Lambda(lambda x: x))
if scale != (0, 0):
if center_norm:
lnorm = CenterNormalizer(scale[0])
out_transforms.append(
transforms.Lambda(lambda x: dewarp(lnorm, x)))
out_transforms.append(transforms.Lambda(lambda x: x.convert(mode)))
else:
out_transforms.append(
transforms.Lambda(
lambda x: _fixed_resize(x, scale, Image.LANCZOS)))
if pad:
out_transforms.append(transforms.Pad((pad, 0), fill=255))
out_transforms.append(transforms.ToTensor())
# invert
out_transforms.append(transforms.Lambda(lambda x: x.max() - x))
out_transforms.append(transforms.Lambda(lambda x: x.permute(*perm)))
return transforms.Compose(out_transforms)
def pil_dewarp(im: Image.Image, lnorm: CenterNormalizer) -> Image.Image:
return dewarp(lnorm, im)
