Hi,

I had just decided to try using a sinc function to remove pixelisation. I did a bit of research googled the mailing lists (well everywhere in fact) but did not find much.

I decided to open and image in Gimp and expand it an try a few values in a convulution matrix. To my frustration it got automatically smoothed.

I soon found the interpelation = none option and tried the other options. I'd not heard of Lanczos Best before so I googled and found out it was the sinc I was trying to do!

Nice work, preemptive development. ;)

It seems pretty good on the tests I did although arguably whether it is better or worse than cubic on my data. Different sure. I think "best" is subjective and depends on the nature of the image.

There was on overall sharpness better than cubic but more aritifacts. This surprised me a bit so I started looking at the code to see if I could determine the parameters and maybe experiment a bit to improve it.

Due to the size of Gimp and the abstraction of all these things going through pdb etc, I did not find all I wanted. However, it did seem that all these resize/rotations etc were operated on a 3x3 matrix and that this dimension is hardcoded.

Could s.o. correct me if I misread that , but it seems to explain why sinc is not giving better results than cubic. If that is the case the complex sinc fn is being represented by only 3 data points (in 1D). While 3 data points may describe a simple cubic interpollation reasonalby well . It in no way represents the complexity of sinc, and so all the F.T. theory behind it falls flat.

My aim is to try a larger matrix for sinc. Could someone point me to where kernel is calculated for sinc and confirm my impression that the resize is fixed to operate on 3x3?

Thanks for any clues. I will save a lot of digging.

regards. GG.