1

I am working on blurred image classification and found that FFT can be applied for blurred image classification. But as I dont have much idea about FFT I have some problems with getting a rotation invariant power spectrum histogram from FFT.

These are some of the codes I tried but I have no idea about how to get a rotation invariant histogram from it. (dont know how to transform this to polar domain and histogram it)

                 fftI = ifftshift(fft2(I));
                 fftI = abs(fftI);
                 to do 1: transform to polar coordinates
                 to do 2: get a rotation invariant histgram with N bins

Can anyone help me? Thanks.

user570593
  • 3,420
  • 12
  • 56
  • 91
  • I recommend this for log-polar sampling: http://www.mathworks.com/matlabcentral/fileexchange/27023-log-polar-image-sampling/content/logsample.m Log-polar will handle rotation and scale in your Fourier spectrum. – Cape Code Apr 29 '14 at 14:49
  • I think DCT is better for this ,... blurred image has less or no energy on higher frequencies so most energy will be on upper left corner of DCT output... if you merge x/y axises together (something like project them on diagonal) then the result is kind of invariant on rotation +/- aspect-ratio and aliasing fragments – Spektre Apr 30 '14 at 08:02
  • @Spektre so will FFT, only that low frequency is in found in the center of a (shifted) 2-D FFT while higher freqencies are closer to the border of the image. But the methode of merging the axis is valid i think – ben Apr 30 '14 at 09:59
  • DFT has tendency to be very picky on phase for each frequency so without proper windowing and filtering there is much more difference then in DCT for comparing two identical but differently rotated or shifted images. At least that is mine experience. – Spektre Apr 30 '14 at 13:12

0 Answers0