research diary

I went into the lab to work on getting what will now be called the pattern matching by cross-correlation algorithm working. It turns out that it really was working all along, but the results looked wrong because I was viewing the absolute value of the results instead of the real components. This should be fine, because the imaginary components are nearly zero, but our pixel values are negative, so this results in an inverted image. In any case, it works now. Now I have to work on a scheme to extract the particle locations from the cross-correlation image.

There are several possibilities for optimization. The small reference image should be rotated before the padding is added, otherwise much time will be wasted on interpolating values known to be zero. Also, I suspect that rotation and fourier transform operations may commute, which could be useful if the padding operation could be done in fourier space as well. Actually, I wouldn't be surprised if the padding operation commuted with the fourier transform as well. On the implementation side, we could use the fftw fourier transform library which is multithreaded, to take advantage of the multiple CPUs available to us.