在区域中切片数组-Python

Question

I have to "divide" an nxm array in regions using a mask input.

For example, suppose I have a 20 x 20 array. My mask is the following ( 5 x 5 ) -- always:

where the numbers represent the regions in which the cells take part. This mask is not an input, that is just an ndarray . This mask just represent how I should slice my 20 x 20 at every 5 x 5 neighborhood.

For example, the first region comprehend the indices:

(0,0), (1,0), (1,1), (2,0), (2,1), (2,2)

For each 5 x 5 neighborhood of my 20 x 20 array, I should return the values that are in each of the 8 regions.

I know how to do that with a "standard code", but I wondering if there is a Pythonic way of do that, possible with a concise code.

As a code example, I could do something like:

def slice_in_regions(data, x_dim, y_dim):
    for x in xrange(0, x_dim, 5):
        for y in xrange(0, y_dim, 5):
            neighbors = data[x:x+4, y:y+4]
            region1 = [neighbors[0,0], neighbors[1,0], neighbors[1,1], neighbors[2,0], neighbors[2,1], neighbors[2,2]]
            # region2, region3...

However, that doesn't seem to be a good way to do that. Moreover, I'm counting on that my data will dimension be multiple of 5 .

Thank you.

Answer 1

It seems you could probably just resize your mask, eg if you're already using numpy ,

mask = mask.repeat(4, axis=0).repeat(4, axis=1)
# Then you apply the mask using 
values = data[mask]

Otherwise,

import numpy as np
mask = np.repeat(mask, 4, axis=0).repeat(4, axis=1)
# Then you apply the mask using 
values = np.array(data)[mask]

Individual regions

If you need to access each region individually, you could precede the previous by using a labelled mask; as the labels will be grown into labelled regions you can then use, eg

values = [ data[mask==l] for l in range(1, mask.max()+1)]

Here values will be a list of arrays where each item corresponds to a labelled region in mask .

Generating the labelled mask

For completeness, to get from a binary mask to a labelled mask where every on pixel has it's own label you could use scipy.ndimage.label

mask = ndimage.label(mask, [[0,0,0],[0,1,0],[0,0,0]])

or if using a region labelling function is overkill, you can achieve a similar result using

mask[mask] = range(1,mask.sum()+1)