Re: Lanczos interpolation

[EMAIL PROTECTED] wrote:
> 
> In rec.photo.digital Guido Vollbeding <[EMAIL PROTECTED]> wrote:
> 
> > If you have JPEG (DCT) data, you can directly use the IDCT for
> > efficient interpolation:
> 
> Ah, that's very cool.  But JPEG tiles don't overlap, do they?  So when
> using this trick for enlarging wouldn't you end up with
> discontinuities at the edges of the tiles?

Yes, but increasing in a reasonable way with increasing enlargement.
That's why I implemented the scaling factors N/8 for all N=1...16.
For N=8 you have the usual JPEG block discontinuities, and then
slightly increasing with N=9...16.
In the mentioned concept, the new (block) edge sampling points lie
*outside* the original (block) edge sampling points when upscaling
(the mentioned 1/(2N) interval at first and last sampling point).
Thus we have rather *extra*polation for these (block) edge points,
and therefore increasing block artifacts.
For downscaling the new sampling points always lie inbetween the
original sampling points - that's why we have real interpolation
and the results are optimal (decreasing block discontinuities).

In any case the interval sampling subdivision is phase-uniform
for all blocks in the Direct (I)DCT method, therefore the
efficient implementation with fixed (I)DCT kernel similar to
the normal (8-point) case.
For *arbitrary* scaling or other image warping applications you
don't have a uniform sampling phase over all blocks, therefore
you can't use a fixed kernel, but it would still be interesting
to see the result and compare the cost with other algorithms.
Quality should be top, except for possibly increasing block
edge artifacts (new sampling points may fall outside original
scope in a block even for downsampling).

Regards
Guido