Invariant Statistics and Coding of Natural Microimages
Don Geman and Alexey Koloydenko
We search for universal characteristics of the microstructure of
natural images. Our data consist of a very large set of 3 by
3 patches randomly extracted from indoor and outdoor grey level
scenes. The patches are grouped into natural equivalence classes
(``patterns'') based on photometry, ``complexity'' and geometry. We
analyze the stability of the pattern statistics over image sets,
resolutions and grey scale distortions. Important aspects of the
probability distribution of the patterns, e.g., the dominant masses,
are stable in our experiments. We also compare the statistics of
the natural patch world with those of artificially generated images;
the results are consistent with recently proposed ``scaling laws''
for the sizes of objects in natural images. These results suggest
that well-chosen patch labels might serve as elementary features in
pattern recognition and other imaging problems in which the fine
structure of the grey level configurations is not essential, and we
sketch a computationally efficient way to carry this out using
tree-structured vector quantization.