Related
I have this 6x6 matrix filled with 0s. I got the secondary diagonal in sec_diag. The thing I am trying to do is to change the values of above the sec_diag inside the matrix with the odds numbers from 9-1 [9,7,5,3,1]
import numpy as np
x = np.zeros((6,6), int)
sec_diag = np.diagonal(np.fliplr(x), offset=1)
The result should look like this:
[[0,0,0,0,9,0],
[0,0,0,7,0,0],
[0,0,5,0,0,0],
[0,3,0,0,0,0],
[1,0,0,0,0,0],
[0,0,0,0,0,0]]
EDIT: np.fill_diagonal isn't going to work.
You should use roll
x = np.zeros((6,6),dtype=np.int32)
np.fill_diagonal(np.fliplr(x), [9,7,5,3,1,0])
xr = np.roll(x,-1,axis=1)
print(xr)
Output
[[0 0 0 0 9 0]
[0 0 0 7 0 0]
[0 0 5 0 0 0]
[0 3 0 0 0 0]
[1 0 0 0 0 0]
[0 0 0 0 0 0]]
Maybe you should try with a double loop
I'm struggling when writing a function that would seemlessly apply to any numpy arrays whatever its dimension.
At one point in my code, I have boolean arrays that I consider as mask for other arrays (0 = not passing, 1 = passing).
I would like to "enlarge" those mask arrays by overriding zeros adjacent to ones on a defined range.
Example :
input = [0,0,0,0,0,1,0,0,0,0,1,0,0,0]
enlarged_by_1 = [0,0,0,0,1,1,1,0,0,1,1,1,0,0]
enlarged_by_2 = [0,0,0,1,1,1,1,1,1,1,1,1,1,0]
input = [[0,0,0,1,0,0,1,0],
[0,1,0,0,0,0,0,0],
[0,0,0,0,0,0,1,0]]
enlarged_by_1 = [[0,0,1,1,1,1,1,1],
[1,1,1,0,0,0,0,0],
[0,0,0,0,0,1,1,1]]
This is pretty straighforward when inputs are 1D.
However, I would like this function to take seemlessy 1D, matrix, 3D, and so on.
So for a matrix, the same logic would be applied to each lines.
I read about ellipsis, but it does not seem to be applicable in my case.
Flattening the input applying the logic and reshaping the array would lead to possible contamination between individual arrays.
I do not want to go through testing the shape of input numpy array / recursive function as it does not seems very clean to me.
Would you have some suggestions ?
The operation that you are described seems very much like a convolution operation followed by clipping to ensure that values remain 0 or 1.
For your example input:
import numpy as np
input = np.array([0,0,0,0,0,1,0,0,0,0,1,0,0,0], dtype=int)
print(input)
def enlarge_ones(x, k):
mask = np.ones(2*k+1, dtype=int)
return np.clip(np.convolve(x, mask, mode='same'), 0, 1).astype(int)
print(enlarge_ones(input, k=1))
print(enlarge_ones(input, k=3))
which yields
[0 0 0 0 0 1 0 0 0 0 1 0 0 0]
[0 0 0 0 1 1 1 0 0 1 1 1 0 0]
[0 0 1 1 1 1 1 1 1 1 1 1 1 1]
numpy.convolve only works for 1-d arrays. However, one can imagine a for loop over the number of array dimensions and another for loop over each array. In other words, for a 2-d matrix first operate on every row and then on every column. You get the idea for nd-array with more dimensions. In other words the enlarge_ones would become something like:
def enlarge_ones(x, k):
n = len(x.shape)
if n == 1:
mask = np.ones(2*k+1, dtype=int)
return np.clip(np.convolve(x, mask, mode='same')[:len(x)], 0, 1).astype(int)
else:
x = x.copy()
for d in range(n):
for i in np.ndindex(x.shape[:-1]):
x[i] = enlarge_ones(x[i], k) # x[i] is 1-d
x = x.transpose(list(range(1, n)) + [0])
return x
Note the use of np.transpose to rotate the dimensions so that np.convolve is applied to the 1-d along each dimension. This is exactly n times, which returns the matrix to original shape at the end.
x = np.zeros((3, 5, 7), dtype=int)
x[1, 2, 2] = 1
print(x)
print(enlarge_ones(x, k=1))
[[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]]
[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 1 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]]
[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]]]
[[[0 0 0 0 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 0 0 0 0 0 0]]
[[0 0 0 0 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 0 0 0 0 0 0]]
[[0 0 0 0 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 1 1 1 0 0 0]
[0 0 0 0 0 0 0]]]
I am looking for the coordinates of connected blobs in a binary image (2d numpy array of 0 or 1).
The skimage library provides a very fast way to label blobs within the array (which I found from similar SO posts). However I want a list of the coordinates of the blob, not a labelled array. I have a solution which extracts the coordinates from the labelled image. But it is very slow. Far slower than the inital labelling.
Minimal Reproducible example:
import timeit
from skimage import measure
import numpy as np
binary_image = np.array([
[0,1,0,0,1,1,0,1,1,0,0,1],
[0,1,0,1,1,1,0,1,1,1,0,1],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,1,1,1,1,0,0,0,0,1,0,0],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,0,1,0,0,0,0,0,0,0,0,0],
[0,1,0,0,1,1,0,1,1,0,0,1],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,1,1,1,1,0,0,0,0,1,0,0],
])
print(f"\n\n2d array of type: {type(binary_image)}:")
print(binary_image)
labels = measure.label(binary_image)
print(f"\n\n2d array with connected blobs labelled of type {type(labels)}:")
print(labels)
def extract_blobs_from_labelled_array(labelled_array):
# The goal is to obtain lists of the coordinates
# Of each distinct blob.
blobs = []
label = 1
while True:
indices_of_label = np.where(labelled_array==label)
if not indices_of_label[0].size > 0:
break
else:
blob =list(zip(*indices_of_label))
label+=1
blobs.append(blob)
if __name__ == "__main__":
print("\n\nBeginning extract_blobs_from_labelled_array timing\n")
print("Time taken:")
print(
timeit.timeit(
'extract_blobs_from_labelled_array(labels)',
globals=globals(),
number=1
)
)
print("\n\n")
Output:
2d array of type: <class 'numpy.ndarray'>:
[[0 1 0 0 1 1 0 1 1 0 0 1]
[0 1 0 1 1 1 0 1 1 1 0 1]
[0 0 0 0 0 0 0 1 1 1 0 0]
[0 1 1 1 1 0 0 0 0 1 0 0]
[0 0 0 0 0 0 0 1 1 1 0 0]
[0 0 1 0 0 0 0 0 0 0 0 0]
[0 1 0 0 1 1 0 1 1 0 0 1]
[0 0 0 0 0 0 0 1 1 1 0 0]
[0 1 1 1 1 0 0 0 0 1 0 0]]
2d array with connected blobs labelled of type <class 'numpy.ndarray'>:
[[ 0 1 0 0 2 2 0 3 3 0 0 4]
[ 0 1 0 2 2 2 0 3 3 3 0 4]
[ 0 0 0 0 0 0 0 3 3 3 0 0]
[ 0 5 5 5 5 0 0 0 0 3 0 0]
[ 0 0 0 0 0 0 0 3 3 3 0 0]
[ 0 0 6 0 0 0 0 0 0 0 0 0]
[ 0 6 0 0 7 7 0 8 8 0 0 9]
[ 0 0 0 0 0 0 0 8 8 8 0 0]
[ 0 10 10 10 10 0 0 0 0 8 0 0]]
Beginning extract_blobs_from_labelled_array timing
Time taken:
9.346099977847189e-05
9e-05 is small but so is this image for the example. In reality I am working with very high resolution images for which the function takes approximately 10 minutes.
Is there a faster way to do this?
Side note: I'm only using list(zip()) to try get the numpy coordinates into something I'm used to (I don't use numpy much just Python). Should I be skipping this and just using the coordinates to index as-is? Will that speed it up?
The part of the code that slow is here:
while True:
indices_of_label = np.where(labelled_array==label)
if not indices_of_label[0].size > 0:
break
else:
blob =list(zip(*indices_of_label))
label+=1
blobs.append(blob)
First, a complete aside: you should avoid using while True when you know the number of elements you will be iterating over. It's a recipe for hard-to-find infinite-loop bugs.
Instead, you should use:
for label in range(np.max(labels)):
and then you can ignore the if ...: break.
A second issue is indeed that you are using list(zip(*)), which is slow compared to NumPy functions. Here you could get approximately the same result with np.transpose(indices_of_label), which will get you a 2D array of shape (n_coords, n_dim), ie (n_coords, 2).
But the Big Issue is the expression labelled_array == label. This will examine every pixel of the image once for every label. (Twice, actually, because then you run np.where(), which takes another pass.) This is a lot of unnecessary work, as the coordinates can be found in one pass.
The scikit-image function skimage.measure.regionprops can do this for you. regionprops goes over the image once and returns a list containing one RegionProps object per label. The object has a .coords attribute containing the coordinates of each pixel in the blob. So, here's your code, modified to use that function:
import timeit
from skimage import measure
import numpy as np
binary_image = np.array([
[0,1,0,0,1,1,0,1,1,0,0,1],
[0,1,0,1,1,1,0,1,1,1,0,1],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,1,1,1,1,0,0,0,0,1,0,0],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,0,1,0,0,0,0,0,0,0,0,0],
[0,1,0,0,1,1,0,1,1,0,0,1],
[0,0,0,0,0,0,0,1,1,1,0,0],
[0,1,1,1,1,0,0,0,0,1,0,0],
])
print(f"\n\n2d array of type: {type(binary_image)}:")
print(binary_image)
labels = measure.label(binary_image)
print(f"\n\n2d array with connected blobs labelled of type {type(labels)}:")
print(labels)
def extract_blobs_from_labelled_array(labelled_array):
"""Return a list containing coordinates of pixels in each blob."""
props = measure.regionprops(labelled_array)
blobs = [p.coords for p in props]
return blobs
if __name__ == "__main__":
print("\n\nBeginning extract_blobs_from_labelled_array timing\n")
print("Time taken:")
print(
timeit.timeit(
'extract_blobs_from_labelled_array(labels)',
globals=globals(),
number=1
)
)
print("\n\n")
My goal is to count objects in a binary array, using Python. I am applying the scikit-image measure.label, to to count objects(should be 1's) in the the array, despite reading the documentation-link, I am getting results that cannot be explained.
a=np.array(np.matrix('0 1 0 0 1;0 1 0 0 0; 0 0 0 0 0;0 0 0 0 1'))
print(a)
img=measure.label(a)
propsa = measure.regionprops(img)
length = len(propsa)
print ('length='+str(length))
for label in propsa:
print (label.centroid)
>>>
[[0 1 0 0 1]
[0 1 0 0 0]
[0 0 0 0 0]
[0 0 0 0 1]]
length=3
(0.5, 1.0)
(0.0, 4.0)
(3.0, 4.0)
When background is selected to be zero,
a=np.array(np.matrix('0 1 0 0 1;0 1 0 0 0; 0 0 0 0 0;0 0 0 0 1'))
print(a)
img=measure.label(a, background=0)
propsa = measure.regionprops(img)
length = len(propsa)
print ('length='+str(length))
for label in propsa:
print (label.centroid)
>>>
[[0 1 0 0 1]
[0 1 0 0 0]
[0 0 0 0 0]
[0 0 0 0 1]]
length=2
(0.0, 4.0)
(3.0, 4.0)
Why is there inconsistent? From my understanding the label function labels "0" as -1 background?!
Following question
It appears that measure.label of the CSV file example counts two objects a large one and its cavity. Sure enough when I inquiry the coordinates of the cavity of img I get a value of 2. Which means the cavity is the second object. Why is an aggregate of zeros counted as an object, and is there away around it?
length=2
(214.23444957510378, 505.25546156532539)
(238.77173913043478, 740.28260869565213)
>>> img[238,740]
2
>>>
For debug purposes it is useful to print the full labeled image. With background=0:
>>> print(img)
[[-1 0 -1 -1 1]
[-1 0 -1 -1 -1]
[-1 -1 -1 -1 -1]
[-1 -1 -1 -1 2]]
The background is correctly labeled as -1. But when you call regionprops on it, it only returns RegionProperties objects for labels 1 and 2 because, as stated in the docs for regionprops(),
label_image : (N, M) ndarray
Labeled input image. Labels with value 0 are ignored.
Therefore the first area that has label 0 is ignored.
When background is not specified, the 0-filled area is has the label 0 and therefore ignored by regionprops(), giving the output of the remaining three 1-filled regions:
>>> print(img)
[[0 1 0 0 2]
[0 1 0 0 0]
[0 0 0 0 0]
[0 0 0 0 3]]
Ok here is simple solution that struck me.
I can simply define background=0, and img=img+1.
The problem was, when label is applied on the matrix with a background=0 the 0 values are changed to -1 and if I have a group of ones they are reduced to 0.Therefore by adding a 1, I adjust the img object to background=0 and any group of numbers that is not 0 will be counted.
Here is an example of what I mean:
import matplotlib
matplotlib.use('Gtk3Agg')
import numpy as np
from skimage import filters, morphology, measure
np.set_printoptions(threshold=np.nan)
a=np.array(np.matrix('0 1 0 0 1;0 1 0 0 0; 0 0 0 0 0;0 0 0 1 1'))
print(a)
img=measure.label(a, background=0)
print('img=')
print (img)
#adjusting +1
img=img+1
print('img+1=')
print (img)
propsa = measure.regionprops(img)
length = len(propsa)
print ('length='+str(length))
for label in propsa:
print (label.centroid)
The code returns the following.
>>>
(4, 5)
[[0 1 0 0 1]
[0 1 0 0 0]
[0 0 0 0 0]
[0 0 0 1 1]]
img=
[[-1 0 -1 -1 1]
[-1 0 -1 -1 -1]
[-1 -1 -1 -1 -1]
[-1 -1 -1 2 2]]
img+1=
[[0 1 0 0 2]
[0 1 0 0 0]
[0 0 0 0 0]
[0 0 0 3 3]]
length=3
(0.5, 1.0)
(0.0, 4.0)
(3.0, 3.5)
Suppose you have a 2D numpy array with some random values and surrounding zeros.
Example "tilted rectangle":
import numpy as np
from skimage import transform
img1 = np.zeros((100,100))
img1[25:75,25:75] = 1.
img2 = transform.rotate(img1, 45)
Now I want to find the smallest bounding rectangle for all the nonzero data. For example:
a = np.where(img2 != 0)
bbox = img2[np.min(a[0]):np.max(a[0])+1, np.min(a[1]):np.max(a[1])+1]
What would be the fastest way to achieve this result? I am sure there is a better way since the np.where function takes quite a time if I am e.g. using 1000x1000 data sets.
Edit: Should also work in 3D...
You can roughly halve the execution time by using np.any to reduce the rows and columns that contain non-zero values to 1D vectors, rather than finding the indices of all non-zero values using np.where:
def bbox1(img):
a = np.where(img != 0)
bbox = np.min(a[0]), np.max(a[0]), np.min(a[1]), np.max(a[1])
return bbox
def bbox2(img):
rows = np.any(img, axis=1)
cols = np.any(img, axis=0)
rmin, rmax = np.where(rows)[0][[0, -1]]
cmin, cmax = np.where(cols)[0][[0, -1]]
return rmin, rmax, cmin, cmax
Some benchmarks:
%timeit bbox1(img2)
10000 loops, best of 3: 63.5 µs per loop
%timeit bbox2(img2)
10000 loops, best of 3: 37.1 µs per loop
Extending this approach to the 3D case just involves performing the reduction along each pair of axes:
def bbox2_3D(img):
r = np.any(img, axis=(1, 2))
c = np.any(img, axis=(0, 2))
z = np.any(img, axis=(0, 1))
rmin, rmax = np.where(r)[0][[0, -1]]
cmin, cmax = np.where(c)[0][[0, -1]]
zmin, zmax = np.where(z)[0][[0, -1]]
return rmin, rmax, cmin, cmax, zmin, zmax
It's easy to generalize this to N dimensions by using itertools.combinations to iterate over each unique combination of axes to perform the reduction over:
import itertools
def bbox2_ND(img):
N = img.ndim
out = []
for ax in itertools.combinations(reversed(range(N)), N - 1):
nonzero = np.any(img, axis=ax)
out.extend(np.where(nonzero)[0][[0, -1]])
return tuple(out)
If you know the coordinates of the corners of the original bounding box, the angle of rotation, and the centre of rotation, you could get the coordinates of the transformed bounding box corners directly by computing the corresponding affine transformation matrix and dotting it with the input coordinates:
def bbox_rotate(bbox_in, angle, centre):
rmin, rmax, cmin, cmax = bbox_in
# bounding box corners in homogeneous coordinates
xyz_in = np.array(([[cmin, cmin, cmax, cmax],
[rmin, rmax, rmin, rmax],
[ 1, 1, 1, 1]]))
# translate centre to origin
cr, cc = centre
cent2ori = np.eye(3)
cent2ori[:2, 2] = -cr, -cc
# rotate about the origin
theta = np.deg2rad(angle)
rmat = np.eye(3)
rmat[:2, :2] = np.array([[ np.cos(theta),-np.sin(theta)],
[ np.sin(theta), np.cos(theta)]])
# translate from origin back to centre
ori2cent = np.eye(3)
ori2cent[:2, 2] = cr, cc
# combine transformations (rightmost matrix is applied first)
xyz_out = ori2cent.dot(rmat).dot(cent2ori).dot(xyz_in)
r, c = xyz_out[:2]
rmin = int(r.min())
rmax = int(r.max())
cmin = int(c.min())
cmax = int(c.max())
return rmin, rmax, cmin, cmax
This works out to be very slightly faster than using np.any for your small example array:
%timeit bbox_rotate([25, 75, 25, 75], 45, (50, 50))
10000 loops, best of 3: 33 µs per loop
However, since the speed of this method is independent of the size of the input array, it can be quite a lot faster for larger arrays.
Extending the transformation approach to 3D is slightly more complicated, in that the rotation now has three different components (one about the x-axis, one about the y-axis and one about the z-axis), but the basic method is the same:
def bbox_rotate_3d(bbox_in, angle_x, angle_y, angle_z, centre):
rmin, rmax, cmin, cmax, zmin, zmax = bbox_in
# bounding box corners in homogeneous coordinates
xyzu_in = np.array(([[cmin, cmin, cmin, cmin, cmax, cmax, cmax, cmax],
[rmin, rmin, rmax, rmax, rmin, rmin, rmax, rmax],
[zmin, zmax, zmin, zmax, zmin, zmax, zmin, zmax],
[ 1, 1, 1, 1, 1, 1, 1, 1]]))
# translate centre to origin
cr, cc, cz = centre
cent2ori = np.eye(4)
cent2ori[:3, 3] = -cr, -cc -cz
# rotation about the x-axis
theta = np.deg2rad(angle_x)
rmat_x = np.eye(4)
rmat_x[1:3, 1:3] = np.array([[ np.cos(theta),-np.sin(theta)],
[ np.sin(theta), np.cos(theta)]])
# rotation about the y-axis
theta = np.deg2rad(angle_y)
rmat_y = np.eye(4)
rmat_y[[0, 0, 2, 2], [0, 2, 0, 2]] = (
np.cos(theta), np.sin(theta), -np.sin(theta), np.cos(theta))
# rotation about the z-axis
theta = np.deg2rad(angle_z)
rmat_z = np.eye(4)
rmat_z[:2, :2] = np.array([[ np.cos(theta),-np.sin(theta)],
[ np.sin(theta), np.cos(theta)]])
# translate from origin back to centre
ori2cent = np.eye(4)
ori2cent[:3, 3] = cr, cc, cz
# combine transformations (rightmost matrix is applied first)
tform = ori2cent.dot(rmat_z).dot(rmat_y).dot(rmat_x).dot(cent2ori)
xyzu_out = tform.dot(xyzu_in)
r, c, z = xyzu_out[:3]
rmin = int(r.min())
rmax = int(r.max())
cmin = int(c.min())
cmax = int(c.max())
zmin = int(z.min())
zmax = int(z.max())
return rmin, rmax, cmin, cmax, zmin, zmax
I've essentially just modified the function above using the rotation matrix expressions from here - I haven't had time to write a test-case yet, so use with caution.
Here is an algorithm to calculate the bounding box for N dimensional arrays,
def get_bounding_box(x):
""" Calculates the bounding box of a ndarray"""
mask = x == 0
bbox = []
all_axis = np.arange(x.ndim)
for kdim in all_axis:
nk_dim = np.delete(all_axis, kdim)
mask_i = mask.all(axis=tuple(nk_dim))
dmask_i = np.diff(mask_i)
idx_i = np.nonzero(dmask_i)[0]
if len(idx_i) != 2:
raise ValueError('Algorithm failed, {} does not have 2 elements!'.format(idx_i))
bbox.append(slice(idx_i[0]+1, idx_i[1]+1))
return bbox
which can be used with 2D, 3D, etc arrays as follows,
In [1]: print((img2!=0).astype(int))
...: bbox = get_bounding_box(img2)
...: print((img2[bbox]!=0).astype(int))
...:
[[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 1 1 1 1 1 1 0 0 0 0 0 0 0]
[0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0]
[0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0]
[0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0]
[0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0]
[0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0]
[0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0]
[0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0]
[0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0]
[0 0 0 0 0 0 0 1 1 1 1 1 1 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]]
[[0 0 0 0 0 0 1 1 0 0 0 0 0 0]
[0 0 0 0 0 1 1 1 1 0 0 0 0 0]
[0 0 0 0 1 1 1 1 1 1 0 0 0 0]
[0 0 0 1 1 1 1 1 1 1 1 0 0 0]
[0 0 1 1 1 1 1 1 1 1 1 1 0 0]
[0 1 1 1 1 1 1 1 1 1 1 1 1 0]
[1 1 1 1 1 1 1 1 1 1 1 1 1 1]
[1 1 1 1 1 1 1 1 1 1 1 1 1 1]
[0 1 1 1 1 1 1 1 1 1 1 1 1 0]
[0 0 1 1 1 1 1 1 1 1 1 1 0 0]
[0 0 0 1 1 1 1 1 1 1 1 0 0 0]
[0 0 0 0 1 1 1 1 1 1 0 0 0 0]
[0 0 0 0 0 1 1 1 1 0 0 0 0 0]
[0 0 0 0 0 0 1 1 0 0 0 0 0 0]]
Although replacing the np.diff and np.nonzero calls by one np.where might be better.
I was able to squeeze out a little more performance by replacing np.where with np.argmax and working on a boolean mask.
def bbox(img):
img = (img > 0)
rows = np.any(img, axis=1)
cols = np.any(img, axis=0)
rmin, rmax = np.argmax(rows), img.shape[0] - 1 - np.argmax(np.flipud(rows))
cmin, cmax = np.argmax(cols), img.shape[1] - 1 - np.argmax(np.flipud(cols))
return rmin, rmax, cmin, cmax
This was about 10µs faster for me than the bbox2 solution above on the same benchmark. There should also be a way to just use the result of argmax to find the non-zero rows and columns, avoiding the extra search done by using np.any, but this may require some tricky indexing that I wasn't able to get working efficiently with simple vectorized code.
I know this post is old and has already been answered, but I believe I've identified an optimized approach for large arrays and arrays loaded as np.memmaps.
I was using ali_m's response that was optimized by Allen Zelener for smaller ndarrays, but this approach turns out to be quite slow for np.memmaps.
Below is my implementation that has extremely similar performance speeds to ali_m's approach approach for arrays that fit in the working memory, but that far outperforms when bounding large arrays or np.memmaps.
import numpy as np
from numba import njit, prange
#njit(parallel=True, nogil=True, cache=True)
def bound(volume):
"""
Bounding function to bound large arrays and np.memmaps
volume: A 3D np.array or np.memmap
"""
mins = np.array(volume.shape)
maxes = np.zeros(3)
for z in prange(volume.shape[0]):
for y in range(volume.shape[1]):
for x in range(volume.shape[2]):
if volume[z,y,x]:
if z < mins[0]:
mins[0] = z
elif z > maxes[0]:
maxes[0] = z
if y < mins[1]:
mins[1] = y
elif y > maxes[1]:
maxes[1] = y
if x < mins[2]:
mins[2] = x
elif x > maxes[2]:
maxes[2] = x
return mins, maxes
My approach is somewhat inefficient in the sense that it just iterates over every point rather than flattening the arrays over specific dimensions. However, I found flattening np.memmaps using np.any() with a dimension argument to be quite slow. I tried using numba to speed up the flattening, but it doesn't support np.any() with arguments. As such, I came to my iterative approach that seems to perform quite well.
On my computer (2019 16" MacBook Pro, 6-core i7, 16 GB 2667 MHz DDR4), I'm able to bound a np.memmap with a shape of (1915, 4948, 3227) in ~33 seconds, as opposed to the ali_m approach that takes around ~250 seconds.
Not sure if anyone will ever see this, but hopefully it helps in the niche cases of needing to bound np.memmaps.