[ptx] Enblend Multiresolution Spline Blender

[Andrew C Mihal]

On Fri, 12 Mar 2004, Kai-Uwe Behrmann wrote:

> Can You explain what the algo does? I would be interessted about.

For the nearest neighbor transform, I don't think there is any hope of 
describing it without a whiteboard. I looked at this paper:

Alsuwaiyel, M. and Gavrilova, M. (2000) "On the Distance Transform of 
Binary Images," CISST'2000: The 2000 International Conference on Imaging 
Science, Systems and Technology, Las Vegas, USA, USA, Vol. I, pp. 83-86

Search for gavrilova on google to find it. The paper is, how should I say, 
full of errors. Even though it is only four pages it took me a few nights 
to figure it out. I'm going to try to implement it - I think it can be 
done with only two passes through the image. This is in comparison to the 
thinning algorithm enblend currently uses - which does 4 passes for each 
iteration!

> Can You explain the therm laplacian pyramids in short words for an non
> scientist (me)?

The idea is that you break down an image into its spatial frequency 
components. The bottom of the pyramid is the same size as the original 
image, but contains only the high-frequency components like sharp edges. 
The next level up is half the size in each dimension, and contains the 
next highest frequency components (not so sharp edges). You keep going up 
the pyramid, and each level is smaller and smaller, and contains coarser 
features. The top level is only 8x8 pixels or so, and has just the most 
basic features in the image. Maybe a few pixels for the sky and a few for 
the ground.

Once you build this pyramid, you can get the original image back by adding 
all of the layers together.

I'll put some more images up on the enblend web page to better describe 
what you see on each level. I need to modify enblend to output these 
intermediate data structures anyway so I can hunt down some ways to save 
memory - it's using too much to stitch anything bigger than 60 megapixels 
or so except on really big machines.

Andrew

---------------------------------
Andrew Mihal
www-cad.eecs.berkeley.edu/~mihal
mihal at eecs.berkeley.edu