I think I've found a good filter this time

[BryantMoore]

Triangle

Code: Select all

$ magick rose: -filter triangle -resize 1000% rose_triangle.png

Lanczos EWA

Code: Select all

$ magick rose: -filter lanczos -distort resize 1000% rose_lanczos_ewa.png

Catrom windowed Jinc EWA - Linear RGB and Sigmoidal Contrast

Code: Select all

$ magick rose: -colorspace rgb +sigmoidal-contrast 8 -define filter:window=catrom -define filter:filter=jinc -distort resize 1000% -sigmoidal-contrast 8 -colorspace srgb rose_jinc-catrom_ewa_rgb-sig8.png

Triangle
http://puu.sh/1dbUi

Lanczos EWA
http://puu.sh/1dbUr

Catrom Jinc EWA
http://puu.sh/1dbUw

[anthony]

Catrom-Jinc!
Another weird one. a negative multiplier (windowing function) to a Jinc function.

Typically window functions are meant to modulate (mute) the weighting function (jinc). As such they typically do not have negatives. But if you say it works... then it perhaps deserves a closer look...

Hmmm...
As this is only used the first lobe of the catrom filter (no negatives), my concern is invalid.

The result is almost identical to a Cylindrical Lanczos (Jinc-Jinc) filter. (except for the sigmoidal aspect).
The images above seem to bear this out.

Try a Sigmoidal Cylindrical Lanczos filter, and you'll probably be hard pressed to see a difference.

[BryantMoore]

I should really learn to check for that before I play with the windowing functions.

[NicolasRobidoux]

It makes sense that using the first lobe of Catmull-Rom as a window function would give something reasonable, which one has no reason to think would be better or worse than Welch, a priori. Welch itself is kind of an arbitrary but reasonable (and "cheap" if one does not use a LUT) way of windowing.
Again, the proof would be in the pudding. One sigmoidized rose enlargement falls more than a bit short of "proof".