Legacy ImageMagick Discussions Archive

Perhaps we should start to classify the filters based on some mathematical keys of correspondence, that comes from the filter curve being generated.

For example a quantities such as... as some initial suggestions...

• rolloff (height at 1 or about that distance),
• blur/sigma (a fit of Gaussian curve to the first lobe? or some other statistical measure),
• ringing (height of first negative lobe)

I don't know if these are sensible suggestions (but I think they may be), but it would help to start qualify what makes a specific 'good' filter and how they relate to each other. You may in fact produce a data cloud for the filters that are found to be good for specific purposes (or test images).

If it results in some type of correspondence between 2, 3, or 4 lobe filters, at least for specific images, then we will have verification that we have a good measure for evaluating filters.

In a way, there are three variables, given that we (sorry, I) always use Sinc in tensor and Jinc in EWA as the "low pass component":
-number of lobes
-(de)blur <- not used, generally, with tensor windowed Sinc methods
-windowing function
Classifying EWA windowed-Jinc, even without the above complication, is much more complicated than classifying tensor windowed-Sinc.
-----
And then there is fitting the other methods in there, somehow.

Note I am talking more about the actual final resulting weighting curve of the filter, rather than how that curve was actually created (cubic, windowing, Lagrange).

Basically taking a step back from 'implementation' to try classifying the actual filter effects, and finding how they correspond to image processing results.

It is more of an idea. and one that could form a new PHD thesis. Shame you are no longer at the university.

Anthony:
This would be a worthwhile thing to put together, but my time is too short to actively participate.
Also: In the classification, you need to split "upsampling" from "resampling" from "downsampling" IMHO whenever you start giving "scores".

Pretty good first pass: http://svn.int64.org/viewvc/int64/resam ... rnels.html