def blocks(data, size):
"""A generator yielding sliding bounding boxes with a given size over given data.
Args:
data : provides a shape attribute. E.g. a hdf5 dataset or a (py)tiff image
size : sets the size of the generated bounding boxes.
Returns:
BoundingBox : A bounding box from the given data.
"""
size = np.array(size)
shape = np.array(data.shape)
n_rows, n_cols = np.ceil(shape[:2] / size).astype(int)
for r in range(n_rows):
for c in range(n_cols):
# swap row and column, x = column, y = row
upper_left = (c * size[1], r * size[0])
lower_right = (min((c + 1) * size[1],
shape[1]), min((r + 1) * size[0], shape[0]))
yield BoundingBox.from_corners(upper_left, lower_right)
def crop_from_array(arr,
bbox,
arr_spacing=1.,
wanted_spacing=1.,
interpolation=CUBIC,
dtype=None,
blocksize=8000):
"""Crop bbox from arr and rescale to wanted_spacing.
Assume arr has spacing arr_spacing. Respect bbox.spacing
Args:
arr (arry-like): image to crop from with spacing arr_spacing.
bbox (BoundingBox): defines the crop area with spacing bbox.spacing.
arr_spacing (float): spacing of arr.
wanted_spacing (float): spacing of the returned crop.
interpolation (int): interpolation used for eventual rescaling.
From skimage.transform.rescale
dtype (numpy dtype): integer dtype of the resulting crop.
Returns:
(np.ndarray): cropped image
"""
arr_spacing = float(arr_spacing)
wanted_spacing = float(wanted_spacing)
if bbox is None:
bbox = BoundingBox.from_corners((0, 0),
(arr.shape[1], arr.shape[0]),
spacing=arr_spacing)
else:
bbox = BoundingBox.copy(bbox)
bbox.set_to_spacing(arr_spacing)
assert bbox.spacing == arr_spacing
output_size = tuple([
int(math.floor(s * arr_spacing / wanted_spacing))
for s in bbox.get_hw(round_values=False)
])
if len(arr.shape) > 2:
# Allow cropping from arrays with more than 2 dimensions
output_size = output_size + arr.shape[2:]
if dtype is None:
result_crop = np.zeros(output_size, dtype=arr.dtype)
else:
result_crop = np.zeros(output_size, dtype=dtype)
if not (blocksize * arr_spacing / wanted_spacing).is_integer():
log.warning(
"blocksize * scale is not a natural number:{}. This may lead to problems in blockwise resizing."
.format(blocksize * arr_spacing / wanted_spacing))
for box in blocks(bbox, (blocksize, blocksize)):
# box is part of bbox, has to be moved by vector of upper left corner
box.set_center(
np.array(box.center) +
np.array(bbox.get_corners(round_values=False)[0]))
output_size = tuple([
int(math.floor(s * (arr_spacing / wanted_spacing)))
for s in box.get_hw(round_values=False)
])
if min(output_size) == 0: continue
#pyextrae.eventandcounters(5000, 20)
crop = np.array(box.crop_from_array(arr))
#pyextrae.eventandcounters(5000, 0)
if arr_spacing != wanted_spacing:
# rescale the crop to fit to output_size
#log.debug("crop_from_array is reshaping from {} to {}".format(crop.shape, output_size))
crop = resize(crop,
output_size,
interpolation=interpolation,
antialias=True)
assert crop.shape[:2] == output_size, "{} {}".format(
crop.shape, output_size)
if dtype is not None:
crop = crop.astype(np.float64) / get_maxintensity(
arr) * dtype_limits(dtype)[1]
crop = crop.astype(dtype)
# move box back to origin and insert crop into result_crop
box.set_center(
np.array(box.center) -
np.array(bbox.get_corners(round_values=False)[0]))
box.rescale(arr_spacing / wanted_spacing)
box.set_hw(box.get_hw()[0], box.get_hw()[1])
box.write_to_array(result_crop, crop)
return result_crop
