It has been shown that basic image statistics are significantly
different for paintings of various content and provenance. Though
such statistics are crude for the purpose of classification, they may
be useful for predicting perceptual judgments such as similarity or
preference, since these statistics are related to efficient coding
strategies in the visual system. To test this notion, we mapped the
similarity space for digitized landscape paintings from a major
university museum by collecting pairwise similarity ratings from
observers. A multidimensional scaling analysis of observer responses
showed strong contributions from basic image statistics relevant to
the visual system, though semantic categories appear to play a larger
role in determining perceived similarity. I will discuss our results
and the prospects for using low level image statistics to perform
high-level perceptual discriminations for artworks.
|
