Image Evolution Using 2D Power Spectra

by Michael Gircys
Supervisor: Brian Ross


This research explores the effectiveness of using power spectral density measures in evolutionary art, and in particular, their use in basic compositional guidance. Existing research has effectively used Fourier power spectral density of spatial frequencies as a metric for image classification and retrieval. We postulate that Fourier decomposition can likewise be effective for guiding image synthesis via genetic programming. Our goals were to improve intuition of these measures, evaluate their utility for characterizing image composition, and adapt them in an evolutionary art environment. Experiments explored factors such as fitness strategies, procedural languages, target images, greyscale and colour, and others. Results are varied; some target images were very effective guides for evolution, while others proved difficult to characterize. Some of the successful results were confirmed by a user survey. We also observed and analysed a previously identified phenomenon of spatial frequency properties resident in uncomfortable art, which could lead to further consideration of visual comfort and aesthetics.

Downloads and publications

Copyright (C) 2018 Michael Gircys.

Back up: