normxcorr2 function - spatial / frequency domain selection

Hi,

I am currently performing cross correlation between various images and know that MATLAB decides whether to use spatial or frequency algorithms based on the image size. I would like to know how large an image has to be for the frequency domain to be used over the spatial domain? I would also like to know if I can force MATLAB to use either the spatial or frequency domain by myself?

Thanks in advance for any answers or insights.

Sam

Answer 1

[Wayne King]

Hi Sam, normxcorr2 computes an estimate of the computation time for spatial domain correlation vs. frequency domain. Whichever is the smallest wins.

You could save normxcorr2.m as a different name and then modify the code so that it always did one or the other. For example, to always use spatial domain, you can modify the xcorr2_fast function inside of normxcorr2 as follows

function cross_corr = xcorr2_fast(T,A)

T_size = size(T);
A_size = size(A);
outsize = A_size + T_size - 1;

% figure out when to use spatial domain vs. freq domain
%conv_time = time_conv2(T_size,A_size); % 1 conv2
%fft_time = 3*time_fft2(outsize); % 2 fft2 + 1 ifft2

%if (conv_time < fft_time)
    cross_corr = conv2(rot90(T,2),A);
%else
%    cross_corr = freqxcorr(T,A,outsize);
%end

To only use the Fourier transform:

function cross_corr = xcorr2_fast(T,A)
T_size = size(T);
A_size = size(A);
outsize = A_size + T_size - 1;

% figure out when to use spatial domain vs. freq domain
%conv_time = time_conv2(T_size,A_size); % 1 conv2
%fft_time = 3*time_fft2(outsize); % 2 fft2 + 1 ifft2

%if (conv_time < fft_time)
%    cross_corr = conv2(rot90(T,2),A);
%else
    cross_corr = freqxcorr(T,A,outsize);
%end

[Sam]

Thanks Wayne, I will certainly give that a try.

[yang wang]

wow, king's answer is wonderful!

[Ravi Singh]

Hi wayne,

Could you tell me if either of this method works greatly when images are rotated ?