Perceived Complexity as Normalized, Integrated, Localized Shannon Entropy

TL;DR

The paper explores how perceived complexity in visual stimuli can be measured using localized Shannon entropy integrated and normalized.

Contribution

The novel approach integrates localized entropy across spatial scales to better capture perceived complexity.

Findings

Normalized, integrated localized entropy at low spatial scales aligns with perceived complexity.

The method reveals aesthetic choices of artists through complexity measurements.

Measurements were validated using natural, urban, and synthetic images.

Abstract

Perceived complexity is a key component of sensory brain function as it indicates the number of resources necessary to process incoming information. A recently proposed measure of perceived complexity defined it as normalized Shannon entropy. However, the proposal used probability distributions estimated from the entire sensory signal at once. Here, we first used synthetically created images and abstract expressionism art to show that using such distributions seemed incompatible with perceived complexity. This incompatibility persisted even if we performed the calculations at different scales, that is, in multiple image resolutions. We then proposed an alternate theory that postulated that perceived complexity arose from the integration of localized Shannon entropy. The outcome of this integration was then normalized to define an index of complexity. We measured this index and integrated Shannon entropy in 704 images obtained from natural and urban settings, painted by well-known artists, or created synthetically. Moreover, we studied the dependence of these measurements on the spatial scale used to measure Localized Shannon Entropy. We found that normalized, integrated, localized Shannon entropy at low spatial scales is consistent with the phenomenology of perceived complexity and illustrates interesting aesthetic choices of different artists.

Figures