Presented at the ECTA conference, Barcelona, 2012. Presents research of unsupervised or autonomous evolutionary art using measures for symmetry, compositional balance and liveliness. Won award for Best Student Paper.

1. Evolving art using measures
for symmetry, compositional
balance and liveliness
Eelco den Heijer
Vrije Universiteit Amsterdam

2. Outline
• Introduction
• Symmetry
• Compositional balance
• Liveliness
• Experiments & Results
• Conclusions
• Future work

3. Introduction
• Unsupervised evolutionary art
• No human in the loop
• Aesthetic Measures
• fitness functions, computational aesthetics
• impact on `style’ of resulting image
• Global Contrast Factor, Ralph/ Ross Bell
curve, Machado/ Cardoso, Benford Law

4. Symmetry & Balance
• Symmetry
• Ubiquitous - it’s everywhere...
• Important in design, architecture
• ‘Hard-wired’ in human visual system?
• It’s role in visual art is not straightforward
• Compositional Balance
• Important in graphic design & visual art
• Processing fluency (Reber)

7. Motivation
• Evo* paper 2011
• Multi-Objective Evolutionary art
• Multiple aesthetic measures
• Some combinations work well, some
combinations do not (same `dimension’,
opposing directions)
• Need for aesthetic measures that work on
other ‘dimensions’ (e.g. symmetry)

8. Research Questions
When using unsupervised evolution (i.e. without a
human in the loop);
1.Is it possible to evolve symmetric images?
2.Is it possible to evolve `balanced’ images?
3.Do the aesthetic measures for symmetry
and balance mix well with other aesthetic
measures?

9. Related work (1)
• Several aesthetic measures in unsupervised
evolutionary art
• Machado & Cardoso (1998)
• Image complexity & Processing complexity
• Matkovic et al (2005)
• Global Constrast Factor
• Ross & Ralph (2006)
• Bell curve
• Several others

10. Related work (2)
• Ngo et al (2000)
• Symmetry in GUI screens
• Bauerly and Liu (2005, 2008)
• Aethetic evaluation of symmetry in web pages

11. Calculating Symmetry
• Select two areas (depending on orientation)
• Mirror the second area (using the proper axis)
• Calculate difference in intensity values between all pixels
in the two areas
• If difference is below 0.05 diff=1, else diff=0

12. Symmetry? Relax...
• Is too much symmetry a good thing?
• Finding a `sweet spot’ for symmetry
• We did not find a value
for this `sweet spot’ in
literature
• we used 0.8 in our
experiments

13. Compositional balance
• Compute visual similarity (or distance) between
image regions
• Stricker & Orengo image distance function
(1995)
• Image is `compressed’ to a feature vector

14. Stricker & Orengo

15. Calculating Balance
• Select two areas (depending on symmetry type)
• Determine feature vector for both areas
• Calculate Stricker & Orengo difference

16. Liveliness
• Using only symmetry would lead to a lot of
monochrome images (since they are perfectly
symmetrical...)
• Same goes for compositional balance; two
halves of a monochrome image have identical
feature vectors
• So, we need additional constraints
• Not only symmetrical, but ‘lively’ too

17. Liveliness: how?
• A simple and naive measure for ‘interestingness’
• Our definition;
interestingness = ‘having a high distribution of
intensity values’
• Calculate entropy of intensity values (x=intensity
value):

18. Experiments
• Unsupervised
• No humans...
• Genetic programming
• `Pixel paradigm’

21. Experiment Setup
1 Symmetry (bilateral) (+ liveliness)
2 Relaxed symmetry (+ liveliness)
3 Compositional balance (+ liveliness)
4 Multi-objective (NSGA-II);
a) Symmetry (all directions)
b) liveliness
c) Global Contrast Factor

22. 1. Bilateral symmetry

23. 2. Bilateral symmetry (relaxed)

24. 3. Compositional Balance

25. 4.Combination (GCF + L + Sym)

29. Conclusions (1)
• Our evolutionary art system has no
difficulty in evolving symmetric images
• Relatively `easy’ aesthetic measure
(rapid fitness progression)

30. Conclusions (2)
• It is possible to control the `amount’ of
symmetry in an unsupervised evolutionary art
system
• Compositional balance
• Images often ‘just‘ symmetrical
• Might need additional ‘penalty’

31. Future work
• Other distance functions for compositional
balance (e.g. based on texture)
• Experiments with symmetry using different
representations (e.g. SVG)
• Good test for compositional balance
measure
• Improve compositional balance measure
• Detect blobs, determine their weight, etc.

32. Thank you!
Images and papers at
http://www.few.vu.nl/~eelco
Questions?
eelcodenheijer@gmail.com