Upsampled image FFT contains NaNs.

I upsampled images and when I take the FFT I get NaN values. How can I prevent NaN values from showing when I take the fourier transform ?

Heres my code to upsample. Upsampling frames in a volume:

ps. ref is 3D

vol = upsample(ref);
pad = GetPad(fftr);
fftr = fftn(vol, pad);
fftt = fftn(tar, pad);
% calculate the power spectrum to remove the magnitude
cp_vol = CrossPowerSpectrum(fftr, fftt);
% go back to the time domain and find the peak
shift_vol = ifftn(cp_vol);

function [res] = CrossPowerSpectrum(A, B)
y = A.*conj(B);
scale = abs(y);
res = y./scale;
end

function [new_dim] = GetPad(vol)
dim = nextpow2(size(ref));
new_dim = 2.^dim;
end

function new_vol = upsample(vol)
m = 2;
[x,y,z] = size(vol);
new_vol = zeros(2*x,2*y,z);
for i = 1 : size(vol,3)
    new_vol(1:m:end,1:m:end,i) = vol(:,:,i);
end
sz = m*m;
H = fspecial('average',[sz sz]);
for i = 1 : size(vol,3)
    new_vol(:,:,i) = conv2(new_vol(:,:,i), H, 'same');
end
end

[Matt J]

But there are no FFT operations in the code that you have posted...??

[raheem mian]

My mistake. Added more code. I see Nan values in shift_vol. I see one NaN in cp_vol. I see a 0 in fftt but no Nans. I think the problem is the 0 in fftt. But I dont have this problem when I dont upsample the image. So I thought maybe I am upsampling incorrectly.

Answer 1

[Matt J]

You have a division by zero in this line,

res = y./scale;

That leads to NaNs in cp_vol and hence also in shift_vol.

[raheem mian]

Yeah I understand that but I need to do that calculation for phase correlation. I was wondering if my upsampling looks correct.

[Matt J]

I don't see anything in your upsample() function that would generate NaNs, though that would be easy to check.