'Magic' kernel for image zoom resampling?

Hi folks,

I've recently generated a solution to an image zooming problem, and
can't find anything quite like it on our good old web or usenet. I'm
trying to figure out if I've stumbled on something that might actually
be useful to other people.

My need was to downsample an image by a factor of 8 in each direction
(i.e. 1 downsampled pixel for every 64 original image pixels), and then
upsample the result back to the original resolution. I understand all
the information theoretical aspects of this process; I was trying to
figure out if there was a kernel that did both tasks well (i.e., the
final result was smooth and without stellations or other strange
artifacts), and also did them fast.

After some truly terrible attempts (even though I thought I had sinc
functions well and truly under my belt 20 years ago), I found that the
following recursive algorithm works amazingly well:

1. To downsample or upsample by a factor of 2, use a kernel

1 3 3 1
3 9 9 3
3 9 9 3
1 3 3 1

placed over every second pixel in each direction. (Total normalization
of 64 is relevant for downsampling. For upsampling, only divide by 16,
because every upsampled pixel gets a contribution from four downsampled
pixels' kernels).

2. To downsample or upsample by a factor of 2^k, perform the above
resampling recursively, k times.

The results for upsampling are astoundingly good, really astoundingly
good, and better than what I could get Photoshop to do. (Some years ago
I did my fair share of 'extreme zooming' in some work on the
photographic evidence of the JFK assassination, so I'm very familiar
with all the strange artifacts that can pop out, and the disappointment
that accompanies it.)

I can upload a few images to my website if need be, to demonstrate what
I am talking about (or, equivalently, point me to some test images and
I'll show you what comes out).

For my particular purpose (8:8 downsampling and upsampling), applying
this 'magic' kernel three times yields a kernel that is only 22 x 22 in
size (if you want to precompute the kernel and apply it once, as I
ultimately want to, rather than actually perfoming the doubling process
three times). (Every time you apply it, double the kernel width or
height and add 2, so 2 x 4 + 2 = 10, and 2 x 10 + 2 = 22.) That's
pretty good, considering that, for 800% zooming, it's already 16 pixels
from nearest pixel to nearest pixel, in each dimension.

Of course, if you wanted to resample by something that is not a power
of 2, then you'd need to use the 'magic' kernel for the nearest power
of 2, and then use something more traditional for the final small
adjustment in resolution. That's no major problem, because getting to
the final result from the nearest power of 2 is never worse than
between a 71% and 141% zoom, and just about any resampling method does
a decent job in this range.

My question is: Is this 'magic kernel' something new, or is this trick
known?

The closest I could find on the net is the 'stair interpolation' trick,
which uses Photoshop's bicubic for successive 110% increases, which is
sort of, kind of, the same idea, but not quite. The other resampling
kernels I could find on the net look much more like what I was trying
to do in the first place, but nothing like what I ended up with.

The 'magic' kernel sort of reminds me of the Fast Fourier Transform,
which also gets all sorts of amazing efficiencies with powers of 2, and
then needs a bit of a (non-critical) fudge if you don't quite have a
power of 2.

Oh, and by the way, I know how I arrived at the 'magic' kernel above
(for another aspect of the project that needed just 2:2 downsampling
and upsampling), and it has some nice mathematical properties (it is
'separable' in the sense that it is the product of an 'x' 1-3-3-1 and a
'y' 1-3-3-1, and its normalization is always a power of 2, which means
everything is integer look-ups with bit-shifts, which makes me
extremely happy), but I have no mathematical proof at all of why it
works so damn well when applied to itself recursively.

Anyhow, thanks in advance for any words of advice. And apologies in
advance if I've rediscovered someone's magic kernel, as is bound to be
the case.

John Costella

_______________________________________________
Dr. John P. Costella
BE(Elec)(Hons) BSc(Hons) PhD(Physics) GradDipEd
Melbourne, Victoria, Australia


assassinationscience.com/johncostella