Decided to threshold them to get a nice binary image.
Took Fast Fourier Transform (fft) in 2D.
Then I obtained and plotted the Log's of the Power Spectrum:
squared the real and imaginary components of the fft (separately)
"imshow'd" the log of the abs() of the log of the Power Spectrum Image.
I next repeated the above steps for one low-pass and one high-pass image:
The high-pass image was obtained by AND'ing its Fourier Transform with
a fftshifted image of a black circle which served as a mask - blocking
out all the low-frequency components of the image.
The high-pass image was then rotated by 30 deg. (and cropped back to original
size) - resulting in a Fourier Transform that was also rotated by 30 deg.
from its normal orientation (perpendicular to all 3 bars).
The low-pass image was obtained in the same way as the high-pass, but the
inverse of the same mask was used instead.
The low-pass image was processed in the same way as the original image,
and then zoomed in - to show more detail where the low-frequency information
had concentrated.
Assumed: Power Spectrum = (real value of each pixel of fftImage)^2 + (imaginary
value of each pixel of fftImage)^2
Approximated: I threw away the imaginary components of images which
I obtained using ifft2.
Weakness: meaningful-looking Fourier-Transform images were obtained
empirically: log(abs(log( fftImage))) was a guess.
I think the major weakness of my solutions is:
Algorithm isn't very useful in current form: Except for obtaining
orientation information, results are obtained only after "tweaking" of
mask sizes and power-spectrum logs.
