Question

I’m checking with you if there is a neat numpy solution to resizing down a 2D numpy array (which is an image) using bilinear filtering?

More specifically, my array has the shape (width, height, 4) (as in a rgba image). The downscaling is also only done on “even” steps: i.e. from (w, h, 4) to (w/2, h/2, 4) to (w/4, h/4, 4) etc.

I’ve browsed around for quite some time now but everyone seems to refer to the scipy/PIL versions of imresize.

I want to minimize the number of dependencies on python packages, hence the numpy only requirement.

I just wanted to check with SO before I go implement it in C++ instead.

Answer 1

I don’t think there is any specific solution in numpy, but you should be able to implement it efficiently without leaving the comfort of python. Correct me if I’m wrong, but when the size of the image is divisible by 2, a bilinear filter is basically the same as averaging 4 pixels of the original image to get 1 pixel of the new one, right? Well, if your image size is a power of two, then the following code:

from __future__ import division
import numpy as np
from PIL import Image

def halve_image(image) :
    rows, cols, planes = image.shape
    image = image.astype('uint16')
    image = image.reshape(rows // 2, 2, cols // 2, 2, planes)
    image = image.sum(axis=3).sum(axis=1)
    return ((image + 2) >> 2).astype('uint8')

def mipmap(image) :
    img = image.copy()
    rows, cols, planes = image.shape
    mipmap = np.zeros((rows, cols * 3 // 2, planes), dtype='uint8')
    mipmap[:, :cols, :] = img
    row = 0
    while rows > 1:
        img = halve_image(img)
        rows = img.shape[0]
        mipmap[row:row + rows, cols:cols + img.shape[1], :] = img
        row += rows
    return mipmap

img = np.asarray(Image.open('lena.png'))
Image.fromarray(mipmap(img)).save('lena_mipmap.png')

Produces this output:

With an original image of 512×512, it runs on my system in:

In [3]: img.shape
Out[3]: (512, 512, 4)

In [4]: %timeit mipmap(img)
10 loops, best of 3: 154 ms per loop

This will not work if an odd length of a side ever comes up, but depending on exactly how you want to handle the downsampling for those cases, you should be able to get rid of a full row (or column) of pixels, reshape your image to (rows // 2, 2, cols // 2, 2, planes), so that img[r, :, c, :, p] is a 2×2 matrix of values to interpolate to get a new pixel value.