Re: Alternatives to bilinear interpolation

Hans-Bernhard Broeker <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>...
> Benji <[EMAIL PROTECTED]> wrote:
> 
> >    Oh, I get it! You say I should be working with a bigger image, in
> > the hopes of preserving quality. That's not a very intuative solution,
> > however, since you're adding another lossy operation.
> 
> Scaling up is not lossy unless you got it horribly wrong.

   It isn't if your definition of scale includes magnifying by
integers only. I don't see how floating-point magnifications could
preserve everything, even when enlarging.

> 
> >    Ok, so this is resampling for transforms. Should I also do this to
> > perform a resize--make it bigger before making it smaller?
> 
> Not unless you fixed your scale-down algorithm first.  As-is, it's
> worse the larger the scale-down factor is, so scaling up and then down
> would get you even worse results than just scaling down already does.

   What do you mean exactly by "fix"? Are you saying my algorithm
needs fixing only because it is bilinear? I'm using bicubic
interpolation now (and found a way to reduce the complexity, such that
the debug version's time has come down from 11 seconds to 3, and 1 in
release) Is that fixed? So should I now upscale to do everything,
including the down-size?