Numpy Two-Dimensional Moving Average

Question

I have a 2d numpy array. I want to take the average value of the n nearest entries to each entry, just like taking a sliding average over a one-dimensional array. What is the cleanest way to do this?

Solidago

How do you define "n nearest entries"? Is this a square region centered on the current entry (eg 5x5)? – Electrokinetics

@PeterGibson That would work fine-- the exact shape of the region doesn't matter for my use case. – Solidago

Answer 1

This is a similar concept to applying a filter to an image.

Fortunately, scipy.ndimage.filters has a bunch of functions to do that. The one you're after is scipy.ndimage.uniform_filter.

Can be used like this:

a
=> 
array([[  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.]])

uniform_filter(a, size=3, mode='constant')
=> 
array([[  1.33333333,   2.33333333,   3.        ,   3.66666667,          2.66666667],
       [  3.66666667,   6.        ,   7.        ,   8.        ,          5.66666667],
       [  7.        ,  11.        ,  12.        ,  13.        ,          9.        ],
       [ 10.33333333,  16.        ,  17.        ,  18.        ,         12.33333333],
       [  8.        ,  12.33333333,  13.        ,  13.66666667,          9.33333333]])

If you want a 5x5 filter, use size=5. The mode option controls how the edges are treated. You didn't specify how you want to handle the edges. In this example, the "constant" mode means it treats each item outside the bounds of the array as a constant value of 0 (0 is the default, which can be overridden).

Hooky

Thanks, but the linked documentation is a little sparse. What would be the exact call I would use if my image is called img and I want to calculate the average of a 5x5 grid around each pixel? – Solidago

It's a bummer you can't specify a different mode for each axis. Or, wait, I guess that wouldn't work. – Cuspid