openCV bicubic imresize creates negative values

Question

I noticed that when downsampling matrices in openCV using the bicubic interpolation I get negative values even though the original matrix was all positive.

I attach the following code as an example:

// Declaration of variables
cv::Mat M, MLinear, MCubic;
double minVal, maxVal;
cv::Point minLoc, maxLoc;
// Create random values in M matrix
M = cv::Mat::ones(1000, 1000, CV_64F);
cv::randu(M, cv::Scalar(0), cv::Scalar(1));
minMaxLoc(M, &minVal, &maxVal, &minLoc, &maxLoc);
// Printout smallest value in M
std::cout << "smallest value in M = "<< minVal << std::endl;

// Resize M to quarter area with bicubic interpolation and store in MCubic
cv::resize(M, MCubic, cv::Size(0, 0), 0.5, 0.5, cv::INTER_CUBIC);
// Printout smallest value in MCubic
minMaxLoc(MCubic, &minVal, &maxVal, &minLoc, &maxLoc);
std::cout << "smallest value in MCubic = " << minVal << std::endl;

// Resize M to quarter area with linear interpolation and store in MLinear
cv::resize(M, MLinear, cv::Size(0, 0), 0.5, 0.5, cv::INTER_LINEAR);
// Printout smallest value in MLinear
minMaxLoc(MLinear, &minVal, &maxVal, &minLoc, &maxLoc);
std::cout << "smallest value in MLinear = " << minVal << std::endl;

I don't understand why this happens. I noticed that if I choose random values between [0,100] the smallest value is after resizing is typically ~-24 vs. -0.24 for the range of [0,1] as in the code above.

As a comparison, in Matlab this doesn't occur (I am aware of a slight difference in weighting schemes as appears here: imresize comparison - Matlab/openCV).

Here's a short Matlab code snippet that saves the smallest value in any of 1000 random downsized matrices (original dimensions of eahc matrix 1000x1000):

currentMinVal = 1e6;
for k=1:1000        
    x = rand(1000);
    x = imresize(x,0.5);
    minVal = min(currentMinVal,min(x(:)));
end

Answer 1

As you can see at this answer the bicubic kernel is not non-negative, therefore, in some cases the negative coefficients may dominate and produce negative outputs.

You should also note that Matlab is using 'Antialiasing' by default, which has an effect on the result:

I = zeros(9);I(5,5)=1; 
imresize(I,[5 5],'bicubic') %// with antialiasing
ans =
     0         0         0         0         0
     0    0.0000   -0.0000   -0.0000         0
     0   -0.0000    0.3055    0.0000         0
     0   -0.0000    0.0000    0.0000         0
     0         0         0         0         0

imresize(I,[5 5],'bicubic','Antialiasing',false) %// without antialiasing
ans =
     0         0         0         0         0
     0    0.0003   -0.0160    0.0003         0
     0   -0.0160    1.0000   -0.0160         0
     0    0.0003   -0.0160    0.0003         0
     0         0         0         0         0