Image Interpolation Software Research

[aruzinsky]

The sample I replied to is impressive.

Thank you. If you would like to see a result on a small test image of your choosing, just post it here.

My algorithm allows a practically unlimited number of regularization constraints which express apriori knowledge of the smoothness characteristics of the image. Thus, there are numerous parameters that can be adjusted and there is a huge lattitude for improvement by incorporating new types of constraints.

Incidentally, there is freeware, Focus Magic, that also performs image enlargement based on regularized deconvolution. Although the results are not as good as mine, it is much faster and the results are better than bicubic interpolation. I have to give Acclaim Software credit - they do good work. You can download it from:

http://focusmagic.com/

 

[Greg Sullivan]

It would have to be very very good for me to use it instead of Genuine Fractals. I really like Genuine Fractals, because it produces such crisp edges.

People who view my enlargements are quite amazed at the sharpness - they don't realise that the sharpness is fake, of course, but that doesn't matter to me at all.

Greg.

 

[Michael Long]

O, 3.3e-4 minutes/pixel (3 color channels) on my 650 MHz. Athlon.

3.3e-4 min/pix X 4.0e+6 pix X 0.65 GHz / 2 GhZ / 60 min/hr = 7.15
hours

This was written completely in Borland C++ v. 5.0. If the inner
routines were written in assembly, I would guess a 3X increase in
speed.

Mmmmmm. Could be even better than that. Depends on whether or not your algorithm can make good use of the multi-byte MMX instructions.

Converting floating point calculations to fixed point routines could also be a big win.

Still, seven hours just means I pick an image and start processing it before I go to bed. Next morning, super image!