Roberto Winter wrote:
I guess the whole point is that Lanczos is not really the 'best'
algorithm for reducing images (as suggested in the scalings dialogs whe
selecting the algorithm). Now, isn't there a problem?...
Yes, there is - and the problem isn't just that Lanczos is not the
"best" algorithm for reducing images, *interpolation* of any kind is
inapproprite when reducing images.
Interpolation allows us to estimate pixel values that "fall in the
cracks" between sample points in the original image - which is a great
help when enlarging; however, when reducing, the "correct" approach is
to perform a weighted average of all source pixels that are "covered" by
the destination pixel.
Cubic interpolation gives a tolerable approximation of the "correct"
method; Lanczos Sinc, it would appear, doesn't.
Another point worth considering is that, while for band-limited
photographic images Lanczos performs very well, for artifically sharp
images it can result in almost JPEG-like ringing artifacts.
So while it produces the "best" results for many images - and is
certainly the method that causes the least amount of blurring when
rotating an image by a few degrees, it's not *always* the best choice.
All the best,
Alastair M. Robinson
Gimp-developer mailing list