To compute the discrete Fourier transform of a grayscale image, just use fft2.
To find the DFT of a colour image, please see the answer to your previous question. If that answer did not help, perhaps you could explain why not.
Why do you want to use Fourier transforms to find features? There are more direct methods, but maybe you want to use the FFT to speed up convolutional pattern matching. If that's the case, program the feature finding using conv2 first, then when it works OK you could substitute this FEX contribution to exploit the FFT.
