Binning multidimensional array in numpy

I have a 4d numpy array (these are stacks of imaging data) and would like to perform mean binning along all but one of the axes.

starting with say

x=np.random.random((3,100,100,100))

I want to apply binning to axes 1,2,3 with bin size 10 and average the values in each bin.

expected result would be an array of shape (3,10,10,10)

I have looked into np.reshape like so:

result=x.reshape(3,-1,10,100,100).mean(axis=1)
result=result.reshape(3,10,-1,10,100).mean(axis=2)

and so on, but this messes up the structure of the image arrays

is there a more straightforward way to do this?

3 answers

  • answered 2021-05-05 11:29 Mostafa Ayaz

    Try this:

    x=np.random.random((3,100,100,100))
    
    x_resized=np.zeros((3,10,10,10))
    
    for i in range(len(x_resized[0])):
        for j in range(len(x_resized[0][0])):
            for k in range(len(x_resized[0][0][0])):
                
                x_resized[0,i,j,k]=np.average(x[0,i*10:i*10+10,j*10:j*10+10,k*10:k*10+10])
                x_resized[1,i,j,k]=np.average(x[1,i*10:i*10+10,j*10:j*10+10,k*10:k*10+10])
                x_resized[2,i,j,k]=np.average(x[2,i*10:i*10+10,j*10:j*10+10,k*10:k*10+10])
    

    which performs the averaging blockwise.

  • answered 2021-05-05 11:34 yan ziselman

    How about this:

    import numpy as np
    import skimage.measure
    
    a = np.arange(36).reshape(6, 6)
    
    b = skimage.measure.block_reduce(a, (2,2), np.mean)
    

    output:

    a = 
    [[ 0  1  2  3  4  5]
     [ 6  7  8  9 10 11]
     [12 13 14 15 16 17]
     [18 19 20 21 22 23]
     [24 25 26 27 28 29]
     [30 31 32 33 34 35]]
    b = 
    [[ 3.5  5.5  7.5]
     [15.5 17.5 19.5]
     [27.5 29.5 31.5]]
    

    But instead of my 2d example, you can do that for a block size of (1, 10, 10, 10) of your data.

  • answered 2021-05-05 12:01 james whalley

    #block size 
    bs = (10,10,10)
    s = 1
    shape = [3,
             x.shape[s+0]//bs[0], bs[0],
             x.shape[s+1]//bs[1], bs[1]
             x.shape[s+2]//bs[2], bs[2]]
    result = x.reshape(*shape).mean(axis = (2,4,6))
    

    Possibly a redundant answer at this point, but if you prefer not to use skimage then this should do the same thing.