def cutout(image_path, trimap_path, cutout_path):
"""
Generate a cutout image from an input image and an input trimap.
This method is using closed-form alpha matting as proposed by :cite:`levin2007closed` and multi-level foreground extraction :cite:`germer2020multilevel`.
Parameters
----------
image_path: str
Path of input image
trimap_path: str
Path of input trimap
cutout_path: str
Path of output cutout image
Example
-------
>>> cutout("../data/lemur.png", "../data/lemur_trimap.png", "lemur_cutout.png")
"""
image = load_image(image_path, "RGB")
trimap = load_image(trimap_path, "GRAY")
if image.shape[:2] != trimap.shape[:2]:
raise ValueError("Input image and trimap must have same size")
alpha = estimate_alpha_cf(image, trimap)
foreground = estimate_foreground_ml(image, alpha)
cutout = stack_images(foreground, alpha)
save_image(cutout_path, cutout)
def alpha_matting_cutout(
img,
mask,
foreground_threshold,
background_threshold,
erode_structure_size,
):
base_size = (1000, 1000)
size = img.size
img.thumbnail(base_size, Image.LANCZOS)
mask = mask.resize(img.size, Image.LANCZOS)
img = np.asarray(img)
mask = np.asarray(mask)
# guess likely foreground/background
is_foreground = mask > foreground_threshold
is_background = mask < background_threshold
# erode foreground/background
structure = None
if erode_structure_size > 0:
structure = np.ones((erode_structure_size, erode_structure_size),
dtype=np.int)
is_foreground = binary_erosion(is_foreground, structure=structure)
is_background = binary_erosion(is_background,
structure=structure,
border_value=1)
# build trimap
# 0 = background
# 128 = unknown
# 255 = foreground
trimap = np.full(mask.shape, dtype=np.uint8, fill_value=128)
trimap[is_foreground] = 255
trimap[is_background] = 0
# build the cutout image
img_normalized = img / 255.0
trimap_normalized = trimap / 255.0
alpha = estimate_alpha_cf(img_normalized, trimap_normalized)
foreground = estimate_foreground_ml(img_normalized, alpha)
cutout = stack_images(foreground, alpha)
cutout = np.clip(cutout * 255, 0, 255).astype(np.uint8)
cutout = Image.fromarray(cutout)
cutout = cutout.resize(size, Image.LANCZOS)
return cutout
def alpha_matting_cutout(
self,
foreground_threshold=240,
background_threshold=10,
erode_structure_size=10,
) -> Image:
base_size = (1000, 1000)
size = self.__image.size
self.__image.thumbnail(base_size, Image.LANCZOS)
mask = self.__mask.resize(self.__image.size, Image.LANCZOS)
img = numpy.asarray(self.__image)
mask = numpy.asarray(mask)
self.__image.close()
self.__mask.close()
# Guess likely foreground/background
is_foreground = mask > foreground_threshold
is_background = mask < background_threshold
# Erode foreground/background
structure = None
if erode_structure_size > 0:
structure = numpy.ones(
(erode_structure_size, erode_structure_size), dtype=numpy.int)
is_foreground = binary_erosion(is_foreground, structure=structure)
is_background = binary_erosion(is_background,
structure=structure,
border_value=1)
trimap = numpy.full(mask.shape, dtype=numpy.uint8, fill_value=128)
trimap[is_foreground] = 255
trimap[is_background] = 0
# Build the cutout image
img_normalized = img / 255.0
trimap_normalized = trimap / 255.0
alpha = estimate_alpha_cf(img_normalized, trimap_normalized)
foreground = estimate_foreground_ml(img_normalized, alpha)
cutout = stack_images(foreground, alpha)
cutout = numpy.clip(cutout * 255, 0, 255).astype(numpy.uint8)
cutout = Image.fromarray(cutout)
cutout = cutout.resize(size, Image.LANCZOS)
return cutout
