"""
Numpy implementation of alpha blending based on information in [SO25182421]_
and [WikiAlphaBlend]_.
For more types of blending see [PypiBlendModes]_ and [WikiAlphaBlend]_.
References:
.. [SO25182421] http://stackoverflow.com/questions/25182421/overlay-numpy-alpha
.. [WikiAlphaBlend] https://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending
.. [WikiBlendModes] https://en.wikipedia.org/wiki/Blend_modes
.. [PypiBlendModes] https://pypi.org/project/blend-modes/
"""
import numpy as np
from . import im_core
[docs]
def overlay_alpha_layers(layers, keepalpha=True, dtype=np.float32):
"""
Stacks a sequences of layers on top of one another. The first item is the
topmost layer and the last item is the bottommost layer.
Args:
layers (Sequence[ndarray]): stack of images
keepalpha (bool): if False, the alpha channel is removed after blending
dtype (np.dtype): format for blending computation (defaults to float32)
Returns:
ndarray: raster: the blended images
Example:
>>> import kwimage
>>> keys = ['astro', 'carl', 'stars']
>>> layers = [kwimage.grab_test_image(k, dsize=(100, 100)) for k in keys]
>>> layers = [kwimage.ensure_alpha_channel(g, alpha=.5) for g in layers]
>>> stacked = kwimage.overlay_alpha_layers(layers)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(stacked)
>>> kwplot.show_if_requested()
"""
layer_iter = iter(layers)
img1 = next(layer_iter)
rgb1, alpha1 = _prep_rgb_alpha(img1, dtype=dtype)
for img2 in layer_iter:
rgb2, alpha2 = _prep_rgb_alpha(img2, dtype=dtype)
rgb1, alpha1 = _alpha_blend_inplace(rgb1, alpha1, rgb2, alpha2)
if keepalpha:
raster = np.dstack([rgb1, alpha1[..., None]])
else:
raster = rgb1
return raster
[docs]
def overlay_alpha_images(img1, img2, keepalpha=True, dtype=np.float32,
impl='inplace'):
"""
Places img1 on top of img2 respecting alpha channels.
Works like the Photoshop layers with opacity.
Args:
img1 (ndarray): top image to overlay over img2
img2 (ndarray): base image to superimpose on
keepalpha (bool): if False, the alpha channel is removed after blending
dtype (np.dtype): format for blending computation (defaults to float32)
impl (str): code specifying the backend implementation
Returns:
ndarray: raster: the blended images
TODO:
- [ ] Make fast C++ version of this function
Example:
>>> import kwimage
>>> img1 = kwimage.grab_test_image('astro', dsize=(100, 100))
>>> img2 = kwimage.grab_test_image('carl', dsize=(100, 100))
>>> img1 = kwimage.ensure_alpha_channel(img1, alpha=.5)
>>> img3 = kwimage.overlay_alpha_images(img1, img2)
>>> # xdoctest: +REQUIRES(--show)
>>> import kwplot
>>> kwplot.autompl()
>>> kwplot.imshow(img3)
>>> kwplot.show_if_requested()
Ignore:
import numpy as np
import kwimage
poly = kwimage.Polygon.random().scale((10, 10))
img2 = np.zeros((10, 10, 4))
img1 = np.zeros((10, 10))
indicator = poly.fill(img1)
to_overlay = np.zeros((10, 10) + (4,), dtype=np.float32)
to_overlay = kwimage.Mask(indicator, format='c_mask').draw_on(to_overlay, color='lime')
to_overlay = kwimage.ensure_alpha_channel(to_overlay)
to_overlay[..., 3] = (indicator > 0).astype(np.float32) * 0.5
raster = kwimage.overlay_alpha_images(to_overlay, img2)
# xdoctest: +REQUIRES(--show)
import kwplot
kwplot.autompl()
kwplot.imshow(raster)
kwplot.show_if_requested()
"""
rgb1, alpha1 = _prep_rgb_alpha(img1, dtype=dtype)
rgb2, alpha2 = _prep_rgb_alpha(img2, dtype=dtype)
# Perform the core alpha blending algorithm
if impl == 'simple':
rgb3, alpha3 = _alpha_blend_simple(rgb1, alpha1, rgb2, alpha2)
elif impl == 'inplace':
rgb3, alpha3 = _alpha_blend_inplace(rgb1, alpha1, rgb2, alpha2)
elif impl == 'numexpr1':
rgb3, alpha3 = _alpha_blend_numexpr1(rgb1, alpha1, rgb2, alpha2)
elif impl == 'numexpr2':
rgb3, alpha3 = _alpha_blend_numexpr2(rgb1, alpha1, rgb2, alpha2)
else:
raise ValueError('unknown impl={}'.format(impl))
if keepalpha:
raster = np.dstack([rgb3, alpha3[..., None]])
# Note: if we want to output a 255 img we could do something like this
# out = np.zeros_like(img1)
# out[..., :3] = rgb3
# out[..., 3] = alpha3
else:
raster = rgb3
return raster
[docs]
def _prep_rgb_alpha(img, dtype=np.float32):
img = im_core.ensure_float01(img, dtype=dtype, copy=False)
img = im_core.atleast_3channels(img, copy=False)
c = im_core.num_channels(img)
if c == 4:
# rgb = np.ascontiguousarray(img[..., 0:3])
# alpha = np.ascontiguousarray(img[..., 3])
rgb = img[..., 0:3]
alpha = img[..., 3]
else:
rgb = img
alpha = np.ones_like(img[..., 0])
return rgb, alpha
[docs]
def _alpha_blend_simple(rgb1, alpha1, rgb2, alpha2):
"""
Core alpha blending algorithm
SeeAlso:
_alpha_blend_inplace - alternative implementation
"""
c_alpha1 = (1.0 - alpha1)
alpha3 = alpha1 + alpha2 * c_alpha1
numer1 = (rgb1 * alpha1[..., None])
numer2 = (rgb2 * (alpha2 * c_alpha1)[..., None])
with np.errstate(invalid='ignore'):
rgb3 = (numer1 + numer2) / alpha3[..., None]
rgb3[alpha3 == 0] = 0
return rgb3, alpha3
[docs]
def _alpha_blend_inplace(rgb1, alpha1, rgb2, alpha2):
"""
Uglier but faster(? maybe not) version of the core alpha blending algorithm
using preallocation and in-place computation where possible.
SeeAlso:
_alpha_blend_simple - alternative implementation
Example:
>>> rng = np.random.RandomState(0)
>>> rgb1, rgb2 = rng.rand(10, 10, 3), rng.rand(10, 10, 3)
>>> alpha1, alpha2 = rng.rand(10, 10), rng.rand(10, 10)
>>> f1, f2 = _alpha_blend_inplace(rgb1, alpha1, rgb2, alpha2)
>>> s1, s2 = _alpha_blend_simple(rgb1, alpha1, rgb2, alpha2)
>>> assert np.all(f1 == s1) and np.all(f2 == s2)
>>> alpha1, alpha2 = np.zeros((10, 10)), np.zeros((10, 10))
>>> f1, f2 = _alpha_blend_inplace(rgb1, alpha1, rgb2, alpha2)
>>> s1, s2 = _alpha_blend_simple(rgb1, alpha1, rgb2, alpha2)
>>> assert np.all(f1 == s1) and np.all(f2 == s2)
"""
rgb3 = np.empty_like(rgb1)
temp_rgb = np.empty_like(rgb1)
alpha3 = np.empty_like(alpha1)
temp_alpha = np.empty_like(alpha1)
# hold (1 - alpha1)
np.subtract(1, alpha1, out=temp_alpha)
# alpha3
np.copyto(dst=alpha3, src=temp_alpha)
np.multiply(alpha2, alpha3, out=alpha3)
np.add(alpha1, alpha3, out=alpha3)
# numer1
np.multiply(rgb1, alpha1[..., None], out=rgb3)
# numer2
np.multiply(alpha2, temp_alpha, out=temp_alpha)
np.multiply(rgb2, temp_alpha[..., None], out=temp_rgb)
# (numer1 + numer2)
np.add(rgb3, temp_rgb, out=rgb3)
# removing errstate is actually a significant speedup
with np.errstate(invalid='ignore'):
np.divide(rgb3, alpha3[..., None], out=rgb3)
if not np.all(alpha3):
rgb3[alpha3 == 0] = 0
return rgb3, alpha3
[docs]
def _alpha_blend_numexpr1(rgb1, alpha1, rgb2, alpha2):
""" Alternative. Not well optimized """
import numexpr
alpha1_ = alpha1[..., None] # NOQA
alpha2_ = alpha2[..., None] # NOQA
alpha3 = numexpr.evaluate('alpha1 + alpha2 * (1.0 - alpha1)')
alpha3_ = alpha3[..., None] # NOQA
rgb3 = numexpr.evaluate('((rgb1 * alpha1_) + (rgb2 * alpha2_ * (1.0 - alpha1_))) / alpha3_')
rgb3[alpha3 == 0] = 0
[docs]
def _alpha_blend_numexpr2(rgb1, alpha1, rgb2, alpha2):
""" Alternative. Not well optimized """
import numexpr
c_alpha1 = numexpr.evaluate('1.0 - alpha1')
alpha3 = numexpr.evaluate('alpha1 + alpha2 * c_alpha1')
c_alpha1_ = c_alpha1[..., None] # NOQA
alpha1_ = alpha1[..., None] # NOQA
alpha2_ = alpha2[..., None] # NOQA
alpha3_ = alpha3[..., None] # NOQA
numer1 = numexpr.evaluate('rgb1 * alpha1_') # NOQA
numer2 = numexpr.evaluate('rgb2 * (alpha2_ * c_alpha1_)') # NOQA
with np.errstate(invalid='ignore'):
rgb3 = numexpr.evaluate('(numer1 + numer2) / alpha3_')
rgb3[alpha3 == 0] = 0
return rgb3, alpha3
[docs]
def ensure_alpha_channel(img, alpha=1.0, dtype=np.float32, copy=False):
"""
Returns the input image with 4 channels.
Args:
img (ndarray):
an image with shape [H, W], [H, W, 1], [H, W, 3], or [H, W, 4].
alpha (float | ndarray):
default scalar value for missing alpha channel, or
an ndarray with the same height / width to use explicitly.
dtype (type):
The final output dtype. Should be numpy.float32 or numpy.float64.
copy (bool):
always copy if True, else copy if needed.
Returns:
ndarray: an image with specified dtype with shape [H, W, 4].
Raises:
ValueError - if the input image does not have 1, 3, or 4 input channels
or if the image cannot be converted into a float01 representation
Example:
>>> # Demo with a scalar default alpha value
>>> import kwimage
>>> data0 = np.zeros((5, 5))
>>> data1 = np.zeros((5, 5, 1))
>>> data2 = np.zeros((5, 5, 3))
>>> data3 = np.zeros((5, 5, 4))
>>> ensured0 = kwimage.ensure_alpha_channel(data0, alpha=0.5)
>>> ensured1 = kwimage.ensure_alpha_channel(data1, alpha=0.5)
>>> ensured2 = kwimage.ensure_alpha_channel(data2, alpha=0.5)
>>> ensured3 = kwimage.ensure_alpha_channel(data3, alpha=0.5)
>>> assert np.all(ensured0[..., 3] == 0.5), 'should have been populated'
>>> assert np.all(ensured1[..., 3] == 0.5), 'should have been populated'
>>> assert np.all(ensured2[..., 3] == 0.5), 'should have been populated'
>>> assert np.all(ensured3[..., 3] == 0.0), 'last image already had alpha'
Example:
>>> import kwimage
>>> # Demo with a explicit alpha channel
>>> alpha = np.random.rand(5, 5)
>>> data0 = np.zeros((5, 5))
>>> data1 = np.zeros((5, 5, 1))
>>> data2 = np.zeros((5, 5, 3))
>>> data3 = np.zeros((5, 5, 4))
>>> ensured0 = kwimage.ensure_alpha_channel(data0, alpha=alpha)
>>> ensured1 = kwimage.ensure_alpha_channel(data1, alpha=alpha)
>>> ensured2 = kwimage.ensure_alpha_channel(data2, alpha=alpha)
>>> ensured3 = kwimage.ensure_alpha_channel(data3, alpha=alpha)
>>> assert np.all(ensured0[..., 3] == alpha), 'should have been populated'
>>> assert np.all(ensured1[..., 3] == alpha), 'should have been populated'
>>> assert np.all(ensured2[..., 3] == alpha), 'should have been populated'
>>> assert np.all(ensured3[..., 3] == 0.0), 'last image already had alpha'
"""
img = im_core.ensure_float01(img, dtype=dtype, copy=copy)
c = im_core.num_channels(img)
if c == 4:
return img
else:
if isinstance(alpha, np.ndarray):
alpha_channel = alpha
else:
alpha_channel = np.full(img.shape[0:2], fill_value=alpha, dtype=img.dtype)
if c == 3:
return np.dstack([img, alpha_channel])
elif c == 1:
return np.dstack([img, img, img, alpha_channel])
else:
raise ValueError(
'Cannot ensure alpha. Input image has c={} channels'.format(c))