09-13-08 - 1

All the theory of digital signal processing can get a little silly. People do all this analysis of the aliasing properties and frequency response of these various filters, they come up with these massively complex things like using sinc functions windows by Kaiser-Bessel functions - and then you wind up using it to change an image size by 2X which means you actually only evaluate those functions at 5 discrete values. So it's just a table of 5 floats that looks sort of like a hump.

I'm doing a little lapped image transform. To do a 4x4 DCT you need an 8x8 lap window, but by symmetry and reversibility it's separable to 8 taps, 4 are just mirrored, and half of those are constrained by the mirrored sum = 1 rule, so there are only 2 free floats. So your lap window can be a kaiser or a sin(sin^2) window, or you can just parameterize by two floats and tweak them to maximize PSNR and subjective quality.

No comments:

old rants