imresize(...'bilinear') and interp2(...'linear') return different results

x = [0 0 1 1;...
    0 0 1 1;...
    0 0 0 0;...
    0 0 0 0];
imresize(x, [2 2], 'bilinear');
interp2(x, 1.5:2:3.5, (1.5:2:3.5)', 'linear');

I would expect imresize and interp2 to return the same answer above. They do not, however:

>> imresize(x, [2 2], 'bilinear')

ans =

    0.1094    0.7656
    0.0156    0.1094

>> interp2(x, 1.5:2:3.5, (1.5:2:3.5)', 'linear')

ans =

     0     1
     0     0
     

My assumption is that imresize does not adhere to the common definition of bilinear scaling (a linear weighted sum of the immediate 2 neighours in each dimension). Perhaps it is using the (functionally more sensible, but terminology-wise problematic) definition used by ImageMagick, where a triangular function is used as a continouus prototype, meaning that when used for large scale downsampling, more than 4 input samples are used to form one output sample?

https://en.wikipedia.org/wiki/Bilinear_interpolation

Answer 1

[Rik]

It is a bit hidden away in the documentation, but it has to do with antialiasing when shrinking an image:

x = [0 0 1 1;...
    0 0 1 1;...
    0 0 0 0;...
    0 0 0 0];
imresize(x, [2 2], 'bilinear')

ans = 2×2

0.1094 0.7656 0.0156 0.1094

●

imresize(x, [2 2], 'bilinear','Antialiasing',false)

ans = 2×2

0 1 0 0

●

interp2(x, [1.5 3.5], [1.5 3.5].', 'linear')

ans = 2×2

0 1 0 0

●