MATLAB vs C++ vs OpenCV - imresize

Question

I have the following MATLAB code which I want to transport into C++

Assume Gr is 2d matrix and 1/newscale == 0.5

Gr = imresize(Gr, 1 / newScale);

in the MATLAB documentation:

B = imresize(A, scale) returns image B that is scale times the size of A. The input image A can be a grayscale, RGB, or binary image. If scale is between 0 and 1.0, B is smaller than A. If scale is greater than 1.0, B is larger than A.

So this means I will get a 2D matrix == matrix_width/2 and matrix_height/2
How do I calculate the values? The default according to the docs are coming from cubic interpolation for nearest 4X4.

I can't find a sample code for C++ that does the same. Can you please provide a link to such code?

I also found this OpenCV function, resize.

Does it do the same as the MATLAB one?

Answer 1

Yes, just be aware that MATLAB's imresize has anti-aliasing enabled by default:

imresize(A,scale,'bilinear')

vs. what you would get with cv::resize(), which does not have anti-aliasing:

imresize(A,scale,'bilinear','AntiAliasing',false)

And as Amro mentioned, the default in MATLAB is bicubic, so be sure to specify.

Bilinear

No code modifications are necessary to get matching results with bilinear interpolation.

Example OpenCV snippet:

cv::Mat src(4, 4, CV_32F);
for (int i = 0; i < 16; ++i)
    src.at<float>(i) = i;

std::cout << src << std::endl;

cv::Mat dst;
cv::resize(src, dst, Size(0, 0), 0.5, 0.5, INTER_LINEAR);

std::cout << dst << std::endl;

Output (OpenCV)

[0, 1, 2, 3;
  4, 5, 6, 7;
  8, 9, 10, 11;
  12, 13, 14, 15]

[2.5, 4.5;
  10.5, 12.5]

MATLAB

>> M = reshape(0:15,4,4).';
>> imresize(M,0.5,'bilinear','AntiAliasing',true)
ans =
                     3.125                     4.875
                    10.125                    11.875
>> imresize(M,0.5,'bilinear','AntiAliasing',false)
ans =
                       2.5                       4.5
                      10.5                      12.5

Note that the results are the same with anti-aliasing turned off.

Bicubic Difference

However, between 'bicubic' and INTER_CUBIC, the results are different on account of the weighting scheme! See here for details on the mathematical difference. The issue is in the interpolateCubic() function that computes the cubic interpolant's coefficients, where a constant of a = -0.75 is used rather than a = -0.5 like in MATLAB. However, if you edit imgwarp.cpp and change the code :

static inline void interpolateCubic( float x, float* coeffs )
{
    const float A = -0.75f;
    ...

to:

static inline void interpolateCubic( float x, float* coeffs )
{
    const float A = -0.50f;
    ...

and rebuild OpenCV (tip: disable CUDA and the gpu module for short compile time), then you get the same results:

MATLAB

>> imresize(M,0.5,'bicubic','AntiAliasing',false)
ans =
                    2.1875                    4.3125
                   10.6875                   12.8125

OpenCV

[0, 1, 2, 3;
  4, 5, 6, 7;
  8, 9, 10, 11;
  12, 13, 14, 15]
[2.1875, 4.3125;
  10.6875, 12.8125]

More about cubic HERE.

Answer 2

In OpenCV, the call would be:

cv::Mat dst;
cv::resize(src, dst, Size(0,0), 0.5, 0.5, INTER_CUBIC);

You might then have to do some smoothing/blurring to emulate the anti-aliasing which MATLAB also performs by default (see @chappjc's answer)