Question

im using the following code to calculate convolution of an image with a specified kernel(in my case gaussian). Everytime I get a different result and the result image is not even close to the one i obtained by convolution in the Spatial domain. First I thought the problem is with the datatype of the images. I changed them to 32 and 64 but still the same results. Can anyone tell me what could be wrong?

http://opencv.willowgarage.com/documentation/cpp/core_operations_on_arrays.html#dft
this function above is giving me a black image. I have input in GRAYSCALE.

void convol_fft(const Mat& A,const vector<vector<float>>& kernel2d,Mat& result)
{

    Mat B = Mat(3,3,CV_64F);
    for (int row = 0; row < kernel2d.size(); row++)
        for (int col = 0; col < kernel2d[row].size(); col++){
            B.at<uchar>(row,col) = (uchar)kernel2d[row][col];
        }

    int dft_M = getOptimalDFTSize( A.rows+B.rows-1 );
    int dft_N = getOptimalDFTSize( A.cols+B.cols-1 );
    Mat dft_A = Mat::zeros(dft_M, dft_N, CV_64F);
    Mat dft_B = Mat::zeros(dft_M, dft_N, CV_64F);

    Mat dft_A_part = dft_A(Rect(0, 0, A.cols,A.rows));
    A.convertTo(dft_A_part, dft_A_part.type(), 1, -mean(A)[0]);
    Mat dft_B_part = dft_B(Rect(0, 0, B.cols,B.rows));
    B.convertTo(dft_B_part, dft_B_part.type(), 1, -mean(B)[0]);

    dft(dft_A, dft_A, 0, A.rows);
    dft(dft_B, dft_B, 0, B.rows);

    // set the last parameter to false to compute convolution instead of correlation
    mulSpectrums( dft_A, dft_B, dft_A, 0, false );
    idft(dft_A, dft_A, DFT_SCALE, A.rows + B.rows - 1 );

    result = dft_A(Rect(0, 0, A.cols + B.cols - 1, A.rows + B.rows - 1));
    normalize(result, result, 0, 1, NORM_MINMAX, result.type());
    pow(result, 3., result);

  //  B ^= Scalar::all(255);

}

Answer 1

I am not sure about OpenCV…but this looks suspicious.

for (int row = 0; row < kernel2d.size(); row++)
    for (int col = 0; col < kernel2d[row].size(); col++){
        B.at<uchar>(row,col) = (uchar)kernel2d[row][col];
 }

If you are filling up the B kernel then the row should be kernel2d[col].size(). It looks like you are overrunning the B kernel. What is value of kernel2d.size() ?

Why not just load the values directly? Saving all the function calls.

For gaussian kernel it should look something like {1,2,1,2,3,2,1,2,1}.