Cite as:
Graham D J, Chandler D M, Field D J, 2005, "How alike are natural scenes and paintings? Characterising the spatial statistical properties of a set of digitised, grey-scale images of painted art" Perception 34 ECVP Abstract Supplement
How alike are natural scenes and paintings? Characterising the spatial statistical properties of a set of digitised, grey-scale images of painted art
D J Graham, D M Chandler, D J Field
Natural scenes share a number of statistical properties including power spectra that are distributed as 1/f 2 ( f = spatial frequency), sparse spatial structure, and similar edge co-occurrence statistics. Painted artworks form an interesting class of images because they are human-created interpretations (and often representations) of the natural world. But, whereas natural scenes comprise a wide range of illuminations and viewing angles, paintings are limited by their smaller range of luminances and viewing distances, their roughly 2-D format, and their typically indoor setting. Nevertheless, paintings have captivated humans for millennia, and statistical similarities in their spatial structure could grant insights into the types of spatial patterns humans find compelling. We investigated the spatial statistics of a large database of digitised paintings from the H F Johnson Museum of Art in Ithaca, NY, and compared them to a set of randomly chosen natural-scene images. A set of randomly chosen, grey-scale images of paintings from the Johnson database--which included a diverse set of paintings of Western and non-Western provenance--was characterised in terms of pixel statistics, power spectra, local operator statistics, and other measures. We found that our set of paintings showed lower skewness and kurtosis than the set of natural scenes, both in its intensity distributions and in its response distributions following convolution with a difference-of-Gaussians operator. The set of painted art images was found to have a typical spatial-frequency power spectrum similar to that of natural scenes. We also used a novel over-complete coding technique to give an estimate of the information content for our set of artworks and our set of natural scenes. For all of our statistical measures, noise whose power is distributed as 1/f 2 and whose pixel intensities were Gaussian-distributed served as a control.
[DJG was supported by NIH EY0153932; DJF and DMC were supported by NGA contract HM 1582-05-C-0007 to DJF.]
These web-based abstracts are provided for ease of seaching and access, but certain aspects (such as as mathematics) may not appear in their optimum form. For the final published version of this abstract, please see
ECVP 2005 Abstract Supplement (complete) size: 2356 Kb
