On Tue, Oct 11, 2005 at 01:38:43AM +0200, cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

Err, a quick summary (perhaps not simple) is that Cubic interpolation uses a third order polynomial to guess at values we don't have pixel samples for. Lanczos is from a family of interpolation schemes called windowed sinc.

The sinc function is defined as sin(x)/x (with it being 1 at x=0). It shows up a lot in signal processing, and can be shown to be "ideal" for interpolation (for some definition of ideal). Since the real world isn't ideal, and all trigonometric functions introduce ringing in a non-ideal world, they window the sinc function to limit how far out the ringing can go.

The sinc interpolation schemes tend to do a visually better job of interpolation at the cost of more processing time (but still less than a generic spline implementation).

I'm glad they added it. I probably won't be personally using it much, since my requirements are elsewhere, but it is a very useful thing to have available.

Incidentally, I know gamasutra had an article on sinc interpolation a few years back that should still be easy to find (I accidentally ran across it again a few weeks ago). Also, http://www.mathworld.com is a great resource for general math information.

Jeff

On Tue, Oct 11, 2005 at 01:38:43AM +0200, cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

i found this on wikipedia:
http://en.wikipedia.org/wiki/Lanczos_approximation

carol

On Tue, Oct 11, 2005 at 01:38:43AM +0200, cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

what is more (or less) in my testing of cubic vs linear, cubic was slower not best. cubic produced a fuzzy image and linear seemed to produce the best resized image.

so if you are testing the different scaling options, i suggest you compare Lanczos with Linear.

carol

Carol Spears wrote:

On Tue, Oct 11, 2005 at 01:38:43AM +0200, cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

what is more (or less) in my testing of cubic vs linear, cubic was slower not best. cubic produced a fuzzy image and linear seemed to produce the best resized image.

This is purely anecdotal, but for what it's worth, a while back I was trying to determine which scaling filter to use in a script to do batch-resizing with ImageMagick. I tested every filter with a couple of my digital photos as samples, in ImageMagick and PaintShopPro (IIRC), and to my eye Lanczos was the best; it was certainly at least *among* the best.

-Anthony DiSante http://encodable.com/
http://nodivisions.com/

cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

Actually, on a more generic level, is there some place (other than the source) where one can look up algorithm details pertaining to things like this and filters? If yes, awesome, if not, awesome, that will be a most informative documentation project to embark on.

Harish --
http://delineate.wahgnube.org

On 10/11/05, Harish Narayanan wrote:

cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

Actually, on a more generic level, is there some place (other than the source) where one can look up algorithm details pertaining to things like this and filters? If yes, awesome, if not, awesome, that will be a most informative documentation project to embark on.

Comparisons of different interpolation methods (in this case not for scaling, but resampling for transformations like rotation.)

http://www.path.unimelb.edu.au/~dersch/interpolator/interpolator.html

/pippin

-- «The future is already here. It's just not very evenly distributed» -- William Gibson http://pippin.gimp.org/ http://ffii.org/

Øyvind Kolås wrote:

Comparisons of different interpolation methods (in this case not for scaling, but resampling for transformations like rotation.)

http://www.path.unimelb.edu.au/~dersch/interpolator/interpolator.html

Thank you, that was very informative.

Harish

-- http://delineate.wahgnube.org/

Hi,
pippin@gimp.org (2005-10-13 at 1256.54 +0200):

On 10/11/05, Harish Narayanan wrote:

cedric GEMY wrote:

Testing 2.3, i can there is new interpolation method called Lanczos. It is described as being better than cubic. Does anyboy know simply :) how it works with the picture ?

Actually, on a more generic level, is there some place (other than the source) where one can look up algorithm details pertaining to things like this and filters? If yes, awesome, if not, awesome, that will be a most informative documentation project to embark on.

Comparisons of different interpolation methods (in this case not for scaling, but resampling for transformations like rotation.)

http://www.path.unimelb.edu.au/~dersch/interpolator/interpolator.html

Some comments about uses:
http://www.softimage.com/Products/Other/Illusion/support/tipsntricks/lancos.htm

Some images comparing lanczos vs nearest for scaling down (and an hybrid system): http://imagebeat.com/mandala/thumbnail.html

GSR