Roberto Michelena wrote:
Actually I always wondered why are the interpolation algorithms not parametric, adjustable per a device model for better precision... Even without a previous device model; for example after reading a chart of 4096 patches, instead of building a table with 4096 gridpoints and a simplistic interpolation between them, use the "in-between" patches for tuning the interpolation algorithm via some parameters, and then build a table with fewer gridpoints (or maybe even the same 4096) but additional data to drive the interpolation.
You would be sort of assuming that there is a "fractal" scale similarity in the device behaviour for this to work. Something like a cellular Neugebauer model could be applied in this way, although you would really need a regular sample grid to do this. Note that cellular Neugebauer will tend to have discontinuities between the cells.
A gridpoint might have, for example, one more byte defining a 'sphere of influence'. Or maybe a vector defining an ellipse of influence centered on the gridpoint. Or maybe a vector plus a linear offset value so that the ellipse is not centered. Etc etc. ; Maybe a single byte defining a 'density of space' at the gridpoint. I'm no matematician, but I guess there are ways to go beyond blind interpolation....
Certainly, but it's all a matter of effort vs. reward. I know from experience that it's easy to throw some new thing into the system, spend lots of time tuning it to hopefully give better results, and end up with inconclusive results (ie. no clear difference, some cases better, other cases worse, peak error up with average down, or the reverse etc.)
Graeme Gill.
