Neuroesthetics seeks to identify the neural basis of aesthetic experience - how does the brain give rise to the perception of beauty? A new paper in Network indicates that artists consistently create works which contain the same statistical properties as natural scenes, even when the objects being depicted do not themselves contain such statistics when photographed.
Redies, Hanisch, Blickhan and Denzler review previous work demonstrating that the "spatial frequencies" of natural scenes (essentially, their spatial complexity) follow a 1/f power spectrum, where increased spatial complexity is increasingly restricted to smaller portions of the scene. This property is sometimes referred to as "scale invariance" and is reflected in a variety of natural data, including human reaction time data, spontaneous fluctuations in brain activity, and perhaps in the connectivity of biological networks - neural and social alike.
Redies et al. show that although human faces do not show scale invariant patterns of spatial frequency in photographs (as determined through Fourier analysis of 5776 different photographs taken under a variety of lighting conditions, facial expressions, and occlusions), they are nonetheless rendered this way in paintings (as determined through Fourier analysis of 447 portraits representing a variety of ages, genders and backgrounds, sampled from the Western cultural tradition across 6 centuries, and 8 different artistic media).
EDIT: To the right you can see a log-log plot of spatial frequency on the x-axis ("cycles per image") and power of those frequencies on the y-axis. What you should note is that portraits and natural scenes track each other very closely, particularly in the higher range of spatial frequencies, whereas the two face-photo databases ("AR" and "Yale") tend to show a steeper negative slope. Negative slopes close to 2 typically indicate scale-invariant properties, but only the portraits and natural scenes show slopes close to this value.
The authors also demonstrated that the influence of gender, beardedness, point of view, artistic technique, and century or country of origin was always small or not statistically significant in terms of the spatial frequency distributions manifest in the portraits. However, portraits uniformly had a spatial frequency distribution more similar to photographs of natural scenes than to photographs of human faces. Redies et al. also conducted a series of control experiments to rule out that these differences might be caused by the trivial influences of variance in gamma radiation in art reproduction, nor the complexity of the backgrounds behind the face images, nor whether the subjects of each photo/portrait were wearing hats or other headdress. All images were normalized to one another with respect to inter-eye distance (since closeup images of objects tend to affect their spatial frequency distributions), and converted to monochrome format to allow for comparisons across media like woodcutting with oil paintings.
The authors' hypothesis is that the creation of art involves resonance between an artist's visual system and the work of art itself. Since our visual systems evolved to perceive natural scenes, perhaps good art takes advantage of this specializing by distorting other images to fit those natural scene statistics. Redies et al are quick to say that art is more than just 1/f power spectra - computer generation of images with these power spectra are not always perceived as beautiful.
One way to make sense of these results is to suppose that an important component of aesthetic experience is ease of processing - so we find more those images more beautiful which more closely correspond to the scenes which our visual systems evolved to perceive. Ease of processing, or "processing fluency" relates to preference ratings of a variety of stimuli, so it seems possible that 1/f power spectra may be over-represented in enduring art precisely because scenes with those spatial frequencies are easier to process with our visual system.
Nifty, thanks for posting this. :)
One concern - as far as I can tell, they only used line drawings of faces for the portraits, so they look, in effect, like high spatial frequency filtered faces.
It would be cool to see the data from paintings of faces, and line drawings of non-faces.