Abstract
Art images and natural scenes share a common mathematical property: their radially averaged Fourier spectral power follows a power-law (1/f² characteristics), indicating scale-invariant properties. This study extends these findings to other man-made categories, including cartoons, comics, and manga. Through 2D power spectra analysis and principal component analysis (PCA), we demonstrate that while these images share the power-law distribution of natural scenes, they exhibit significantly higher "slope isotropy"—a more uniform distribution of structural complexity across all orientations. These results suggest that man-made aesthetic stimuli are specifically crafted to evoke pleasant visual perception by mirroring natural statistics while optimizing directional uniformity.
Introduction: The Mathematical Fingerprint of Aesthetics
What defines an image as "aesthetic"? While art history often focuses on cultural context and stylistic evolution, the field of neuroaesthetics seeks universal mathematical properties that resonate with the human visual system. A cornerstone of this research is the concept of scale invariance. In natural scenes, the statistical distribution of structural detail remains constant regardless of magnification—a property represented by a 1/f² power spectrum.
Previous research confirmed that fine art often mimics this natural scaling. However, two major questions remained: does this property extend to popular media like cartoons and manga, and can we distinguish between image categories by looking beyond simple 1D averages to full 2D Fourier spectra?
Methods and Datasets
Researchers analyzed a diverse dataset of 1,500 images, including monochrome fine art, cartoons, Western comics, Japanese manga, and various control groups such as natural scenes, human faces, and scientific illustrations.
The computational pipeline involved:
- Fourier Analysis: Converting images into the frequency domain to quantify "energy" (power) at different spatial frequencies.
- 1D Power Law Fitting: Calculating the radially averaged slope to verify 1/f² characteristics.
- 2D PCA: Using the first 15 principal components of 2D power spectra to determine if image categories occupy distinct statistical spaces.
- Anisotropy Measurement: Dividing the 2D spectra into 32 sectors to measure how mean power and slope vary across different angles (e.g., horizontal vs. vertical vs. oblique).
Key Results: The Universal Aesthetic Slope
The study found that cartoons, comics, and manga strictly adhere to the 1/f² rule, much like fine art and natural scenes. This suggests that even simplified or exaggerated media are "tuned" to the same fractal-like statistics that the human visual system evolved to process efficiently.
As shown in the figure above, the spectral power for these man-made categories decreases linearly on a log-log scale as spatial frequency (cycles per image) increases. The slopes are remarkably consistent, hovering near the -2.0 ideal. In contrast, non-aesthetic categories like scientific illustrations showed shallower slopes (~ -1.5), prioritizing high-frequency edges over naturalistic distribution.
The "Aesthetic Isotropy" Hypothesis
The most significant finding of the research concerns slope isotropy. While both art and nature tend to have more "power" in cardinal orientations (horizontal and vertical lines), art is significantly more uniform in its slope across all angles.
In natural photographs, the "texture" or complexity of the image often changes depending on the direction of observation. In contrast, aesthetic man-made images exhibit higher isotropy; the artist appears to subconsciously balance structural complexity so it remains consistent regardless of orientation. This property held true even when comparing landscape paintings to landscape photographs, suggesting it is a byproduct of the creative process rather than the subject matter itself.
Conclusion
The results indicate that artists—whether drawing a superhero or painting a masterpiece—act as "statistical engineers." They subconsciously craft images that mirror the scaling of the natural world while "smoothing out" the directional biases found in raw physical reality. While these properties are common to images designed for enjoyment, whether they are sufficient to induce aesthetic pleasure remains a subject for future investigation, particularly regarding the role of phase information and higher-order statistics.
Summary updated