MSc thesis: quantitative comparison of enlargement filters

[NicolasRobidoux]

The actual thesis and associated software are still being revised, but the slides of my student Adam Turcotte's The EXQUIRES (EXtensible QUantitative Image RESampling) test suite: impact of the downsampler, difference metric, test image, resampling ratio and colour space on upsampler rank Masters thesis defense are available at http://exquires.ca/slides/slides.pdf.
(Note that the shown ranks for the EANBQH method are almost certainly wrong. Bug fix time. The issue is a collection of casts in the EANBQH code itself.)
Note: IMHO, being a good reconstructor should not automatically put a method at the top of your list of favorites. So, the rankings should be taken with a very very very very large grain of salt. Accuracy, as measured in a reconstruction test, is not the same as "good looking", and at this point no metric accurately measures similarity as humans see it. Also: What the thesis shows is that the rankings are very sensitive to the experimental set up. Basically, the thesis establishes clearly that its own rankings are not gospel.

[henrywho]

Is madvr "commercially available"?

[NicolasRobidoux]

Is madvr "commercially available"?

It's in alpha, and available (for free): http://forum.doom9.org/showthread.php?t=146228 madVR is closed source.
Note however that apparently sigmoidization does not work so well with the often low quality image content of DVDs etc, or else it does not work well with Y'CbCr. I don't know yet whether there is more to it than that, and what is what. In any case, madVR is not considering sigmoidization for inclusion any more, in part because the main dev has a very effective "Anti Ring" ringing post-processor. Actually Mathias Rauen found a png image which does not look good when enlarged with his implementation. I'll have to look at things very very carefully. Manana.

[NicolasRobidoux]

madVR now uses the same version of EWA LanczosSharp that is built into ImageMagick, and a lot of people are really pleased with it. I recommended the "classic" deblur instead of the one I push in my "Recommendations" because I figured it would interact better to Mathias Rauen's AR (AntiRing) filter.

[NicolasRobidoux]

Cristy, Anthony, Fred and Henry: Maybe you want to scan the hopefully final version of my student's thesis http://web.cs.laurentian.ca/nrobidoux/m ... Thesis.pdf for your last name, and let me know if you have any problem with the way it is invoked. There is still time to change little things.

[NicolasRobidoux]

I just realized that we forgot to thank people who contributed test images. Will fix that ASAP.

[fmw42]

Cristy, Anthony, Fred and Henry: Maybe you want to scan the hopefully final version of my student's thesis http://web.cs.laurentian.ca/nrobidoux/m ... Thesis.pdf for your last name, and let me know if you have any problem with the way it is invoked. There is still time to change little things.

My name looks fine.

Fred

[NicolasRobidoux]

Glenn and Cristy:

Are you OK with being acknowledged like this in the acknowledgements?

Cristy and Glenn Randers-Pehrson also deserve thanks for
tirelessly improving ImageMagick.

Do you prefer something different, or not being thanked?

[magick]

Ok to acknowledge. Your wording is fine.

[anthony]

sorry for the slow reply...

Reference 103. The URL is wrong. It was right some time ago, but it should be referencing the 'filter' page and not the 'resize' page.


I am not certain why I am associated with Catmull-Rom filter in EWA. I never proposed it.

Before Nicholas helps me fix the EWA filter I used that recommended Gaussian Blurring filter for EWA. After it was fixed I basically switched to the Mitchell-Netrelli filter, as it just seemed to work better than anything else I tried. Even up to be extremely good indeed.

I do talk about Catmull-Rom but more as a starting point for Cubic filters, as well as its use as a interpolation filter. It is generally a good starting point, but that is all. I never really recommend it beyond its use as a interpolation function.

[NicolasRobidoux]

I am not certain why I am associated with Catmull-Rom filter in EWA. I never proposed it.

Early on, you showed the graph of its impulse response: http://www.imagemagick.org/Usage/filter ... terpolated. That's why.
You still want my student Adam Turcotte to remove you from the "credits" for this method?
(I don't care either way, and he'll do what I say. I think he's filing this coming Monday, though, so please reply ASAP.)
-----
The links will be fixed. This thesis has been quite some time in the making, so it makes sense that some of the web references be stale.

[NicolasRobidoux]

Anthony:
I removed any mention of your name in relation to EWA Catmull-Rom.
If you want me to put you back in, let me know ASAP.

[anthony]

It is fine. Just thought it strange to have been mentioned for that.

The catrom usage there was just to show artefact differences.

I don't mine my name being used or associated.

[NicolasRobidoux]

Adam's putting back the sentence about you showing the cardinal basis function.

[NicolasRobidoux]

Adam's done \o/
Web site: http://exquires.ca
Thesis (will eventually go into arXiv): http://web.cs.laurentian.ca/nrobidoux/m ... Thesis.pdf