[geeks] Re: [rescue] Spaceballs.

[Dave McGuire]

On December 8, Scott Newell wrote:
> >  The time domain for a single image (take a single-channel image for
> >simplicity) is luminance.  Look at it this way...take an image as a
> >2-d array of integer pixels.  Iterate through the rows and columns
> >sequentially, as they would be stored in memory.  Plot the luminance
> >(the integer pixel value).  Bang, there's your time-domain signal.
> 
> I guess you could do that, but all the lit I've seen on image processing
> uses the two dimensional Fourier transform, which has a double integral
> (continuous space) or double summation (discrete space).  
> 
> What you propose would not be rotational invariant--vertical bands would
> contain significant high frequency energy (as it should--it's a sharp
> edge), since your iteration would see the deltas as sequential samples in
> the time domain.  However, the iteration would screw with horizontal bands,
> 'cause you wouldn't get to a transition until you'd iterated over the
> entire row.  (I'm sorry if that doesn't make sense--I'd almost need to draw
> sketches to explain.)

  Nono, you made perfect sense.  You are correct of course, and as I
probably should have pointed out, my example was an application in
which I was looking for the energy distribution across horizontal rows
of pixels.  One could flip the array indices around and do it the
other way, but even doing both still doesn't result in a 2D FFT.

     -Dave

-- 
Dave McGuire
St. Petersburg, FL