The original game plan was to put off doing the prior art search, and try to guess what thr prior art ought to have been. At best, I'd be completely wrong and have a cool strategy. So I read about semi-related things and coded tools that were likely to become useful. The result was the somewhat incoherent proposal section of this site, which bears not all that much resemblance to my current plans.

Here's a summary of what follows plus what hasn't made it in there yet. There are papers about tiling images with other images (using wavelets coefficients or simpler pixel color distance metrics, leaving the colors as they are or shifting them for a better match). There are papers about queryable image databases (the coolest one is like a plug-in for Adobe Illustrator: you draw an image or scan one in and it scales it to 128x128 and takes the wavelet transform; the database has a table of biggest-40-wavelet-coefficients for its images, quantized all the way down to +1/-1 that it searches for best matches screamingly fast). There are several theses and papers about texture matching with oriented filters/steerable pyramids, which was an area I didn't get into.

I don't know where I should cite the image collection CDs I've gotten from the Harvard U. Grad School of Design library. Here, for now. There was the GSD student work 96-97 collection, Josef Albers's Interaction of Color, "Edinburgh: An Interactive Architectural Guide," "Thinking the Unthinkable House," and the clip art CD, "Art Explosion 40,000" (heh), the last of which was used for all of the images on these pages as of project presentations this May.

Jacobs, Charles E.; Finkelsteing, Adam; Salesin, David H; "Fast Multiresolution Image Querying", SIGGRAPH '95.

With their program, the user could query an image database by sketching the image they wanted (or scanning in a poor copy). The program rescaled the drawing to 128x128, quantized the top 40-60 Haar wavelets coefficients (they cite NR's heapselect()) to +1 or -1, and compared the resulting signature against a database of such signatures.

The algorithm for this image-tiling project is similar. However, Daubechies-4 wavelets were used because a small set of those coefficients is much more expressive than the same number of Haars. Also, their dynamic range is so much larger than that of Haar wavelets (exponentially dropping off(?) vs. all sort of in the 0-256 region) that for a coefficient to be in the top 40 is much more significant. OTOH, the +1/-1 quantization becomes necessary now: whereas Haar coefficients for an image at one resolution are the same as at any resolution (except for high-frequency components that disappear), these change by some inconsistent factor of twoish with resolution change (uhhh, I forget with which rescaling algorithm this was associated), such that coefficients were even re-ordered and some wandered on and off the top40 list. At any rate, including a real magnitude comparison would have been tricky. (OT3H, the average-luminance coefficient, though it usually appears in the top 40, is off by an unknown scaling, and quantized, being positive, always to +1, and therefore useless; so that information is a separate, independently-weightable contribution to the score of a match. 

Finkelstein, A.; Range, M.. 
"Image mosaics"

Electronic Publishing, Artistic Imaging, and Digital Typography. 7th
International Conference on Electronic Publishing, EP'98, Held Jointly
with the 4th International Conference on Raster Imaging and Digital
Typography, RIDT'98 Proceedings, p. xiii+574, 11-22
1998

"We describe a process for creating an image mosaic-a collection of
small images arranged in such a way that when they are seen together
from a distance they suggest a larger image. To visually suggest the
larger form, the small images are arranged to match a large picture as
much as possible, and then their colors are adjusted to better suggest
the overall form. Arrangement of the small images may be either manual
or automatic. Adjustment of the colors in the small image to further
suggest the larger picture is fully automatic and employs a new color
correction scheme that generalizes traditional halftoning (14 Refs.)"

1.BARKER         STACKS / Z250.7.I58 1998                In Library

They take images (off the net, hmm) and tile with them, all different,
or all the same, or not.  If not all the same, they use them randomly
or _sometimes take into account the shape/texture/color of the tiles
-- in one case with wavelets (drat, it's been done) -- just a 1st
approx, largest coefficients as signatures to search, so it apparently
_screams.  Then they draw with H/S straight from the tiles and modify
the V (gently) to match the target better.  They note that a video
version (image or tiles change, other doesn't) and irregular tiles
(optimize lots of things at once) haven't been done as of '98...

--------------------
          AUTHOR :Silvers, Robert, 1968-
           TITLE :Photomosaics : putting pictures in their place /
   4.ROTCH          STACKS / THESIS Thesis MediaA&S 1996    In Library
                    M.S. 

and the movie:

Somerville Public Library-Main Branch
The library is at the corner of Highland Avenue and Walnut St.
CALL NUMBER: 779/SI -- Book -- Available 

Apparently what he's doing towards getting shapes and edges to match
up (having tried and discarded wavelets and other complicated
techniques) is actually minimizing the "distance" between small
regions (tens of pixels?) in the tile and the target.

----------

More ML theses

T. Minka '96 EE MEng, F. Liu '97 MAS PhD

Texture identification-- trying to make fourier transforms
"shiftable".

----------

Freeman/Simoncelli papers

on shiftable transforms-- wavelets are great for approximating the
"shape" of a picture, but not textures -- if you move the texture over
by one pixel, its transform is substantially different.  To do
anything much with texture-matching, I want to look at the steerable
pyramids / oriented filters papers.  For something more
photomosaickey, wavelets might be just right.

--------------------

other citations by the Fink/Range article:
----------
SIGGRAPH 90 95 96 97 

90- Paint by Numbers: Abstract Image Representation (including
Dirichlet domains with solid colors, painting with images of brush
strokes, etc.)

95- Fast Multiresolution Image Querying (matching largest wavelet
coefficients -- they take about 20 or 40 or 60 coeffs, and heavily
quantize them (to +/- 1!!)

95- Artistic Screening (multiweight text, escherlike images, etc.)

96- Painterly Rendering for Animation

//96- Multiresolution Video

97- Processing Images and Video for an Impressionist Effect
(brushstrokes perp to gradients, with temporal coherence, computer
vision to move them along with objects)

         TITLE: Computer graphics.
     PUB. INFO: [New York, Special Interest Group on Computer Graphics,
                  Association for Computing Machinery]
   DESCRIPTION: v. ill. 29 cm.
 LINKING NOTES: SIGGRAPH... conference proceedings number continued in 1993 by:
                  Computer graphics proceedings, annual conference series.
      LOCATION: Loeb Design: Per Current issues on Display.
                  Library currently subscribes to this title.

---------- 
 Mandelbrot, Benoit B. The fractal geometry of nature /
    Cabot Science: QA447.M357 1982

----------
 Close, Chuck, 1940- Chuck Close : recent paintings /
    Fine Arts: FA4267.532.30 Folio
 ====> ......................Not checked out.

----------
1993
           TITLE :Storage and retrieval for image and video databases.
       PUBLISHED :Bellingham, Wash., USA : SPIE, c1993-
   PHYSICAL DESC :v. : ill. ; 28 cm.
          SERIES :Proceedings of SPIE--the International Society for Optical
                  Engineering; 
   2.BARKER         STACKS / TA1635.S767 1993               In Library

Has the QBIC stuff -- not so useful -- 

----------

Golan says see: Dali's Hallucinogenic Toreador, and I says see:
Lincoln in Dali-vision.  (So... whaddaya think?)

----------