Use of GCF Aesthetic Measure in the Evolution of Landscape Designs
Prasad Gade and Paul Walsh
Cork Institute of Technology, Bishopstown, Cork, Ireland
Keywords: Interactive Genetic Algorithm, Computational Aesthetic Measure, Image Contrast, Image Complexity,
Fractal Landscapes, Digital Artefact.
Abstract: This paper explores the use of a global contrast factor (GCF) as an aesthetic measure to aid the generation of
fractal landscapes. In an attempt to auto generation virtual landscapes, we added a global contrast factor as
an aesthetic measure based fitness function to the genetic algorithm (GA). This GA is used to explore a
multi-dimensional parameter space that defines how 3D fractal landscapes are created. Two types of
experiments were conducted using GCF that facilitated fluid evaluation of computationally intensive fitness
evaluation, with preliminary results reported.
1 INTRODUCTION
Computer-generated digital artefacts are often
considered to be genuine works of art. They are used
across a variety of fields, for example in advertising,
games development (Halo, 2001); (Assassin’s
Creed, 2007), as well as in the film industry
(Rhythm and Hues Studio, 1987); (Lightwave,
1993). However, designing virtual artefacts is a time
consuming process that requires highly artistic skills
and knowledge of specialist techniques. Users
generally create these using sophisticated drawing
tools and graphic software (Photoshop, 1990)
(Gimp, 1996). There are also semi-automatic
software tools (Vue, 2012); (Bryce, 2010) available,
which can help designers to create 3D artefacts more
intuitively. However, these require a great deal of
manual input, patience and time. Also, users of these
tools often require extensive training and experience
before they can actually deliver the desired product.
A new field of procedural techniques has
emerged recently based on evolutionary algorithms,
where a computer generates digital artefacts
automatically by allowing users to direct an
algorithm towards the desired output, without
requiring any specialist expertise. Authors (Walsh
and Gade, 2010) have implemented such a technique
for generating landscape designs, primarily of
terrains, using an interactive genetic algorithm
(IGA). An IGA is an extended version of a genetic
algorithm (GA) where the fitness evaluation is done
according to the user’s preferences. Figure 1 gives
the result of our work where a user generated digital
landscape designs using real-world scenery (Alpine,
2011); (Desert, 2011) as a target.
Figure 1: Evolution of a real world using IGA.
There are significant drawbacks to this process, i.e.
user fatigue resulting in loss of interest; patience or
miss-guidance of the system during the evaluation
phase. However, we address this by the use of an
aesthetic measure where the evolution of images is
guided without the need for significant user
involvement. In previous work, (Walsh and Gade,
2011) implemented a Kolmogorov complexity
aesthetic measure (Li, 1997) to generate landscape
designs automatically. The results were encouraging
and have led us to integrate additional aesthetic
measures into our library, hoping to improve the
ability of our algorithms to generate more pleasing
landscape designs for users. We investigate the
utility of using computational aesthetic measures to
design and compose artefacts by testing these
concepts within a 3D virtual world.
In any image processing system, the contrast
attribute can play an important role in defining the
83
Gade P. and Walsh P..
Use of GCF Aesthetic Measure in the Evolution of Landscape Designs.
DOI: 10.5220/0004520800830090
In Proceedings of the 5th International Joint Conference on Computational Intelligence (ECTA-2013), pages 83-90
ISBN: 978-989-8565-77-8
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
features of that image. So we implement a new
approach to aesthetic measures in our system, based
on a global contrast factor (GCF) that was
introduced by (Matković et al., 2005) in order to
evolve the best landscape designs with good contrast
levels automatically.
We also attempted to increase the performance
of the IGA system by helping users to identify the
best landscapes in every generation by testing
against the aesthetic fitness scores, GCF and
Kolmogorov complexity. Users can select aesthetic
measures individually, or in combination with each
other, to guide the evolution. In addition, if a user is
not satisfied with the results generated by particular
aesthetic measures, they can choose their own
preferred landscapes by ranking images manually.
Two types of test were conducted in section 6 to
test the effectiveness of the aesthetic measures, GCF
and IGA, in directing a search for evolving
landscapes.
The remainder of the paper is organized as
follows: Section 2 describes previous aesthetic
measures used to evolve digital artefacts; Section 3
describes the GCF aesthetic measure fitness
function; Section 4 describes the parameters
involved in landscape designs; Section 5 explains
the details of the experiment setup; Section 6 shows
the results from the experiments and follows with
conclusions, outlining findings, and future scope.
2 BACKGROUND
(Bentley, 1999) describes how evolutionary designs
generated by computers are surprisingly better than
those designed by humans. They allow the designer
to explore numerous techniques for novel design
concepts. However, as IGAs are not fully
autonomous, researchers implement various
aesthetic measures to help identify the best digital
art that can satisfy human aesthetic tastes, to some
extent.
In recent decades, research has been applied to
the challenge of using evolutionary computation
(EC) with aesthetic measures as fitness scores to
evolve digital artefacts automatically, thus replacing
user involvement. This allows a directed search on a
population of randomly generated individuals over a
number of generations, whereby successive
generations are selected via a fitness function. The
fitness function is a key aspect of this search
heuristic and is commonly based on a computational
measure within the domain of interest, which in this
case is the quality and utility of the generated
landscapes. There are some common aesthetic
measures which are applied to the evaluation of
digital artefacts.
Based on observation of visual preferences on
images selected by humans, authors (Li and Hu,
2010) have selected a set of multiple aesthetic
measures: Bell Curve, (Ross et al., 2006), Image
Complexity Theory (Machado and Cardoso, 1998),
Birkhoff (Birkhoff, 1933) & Shannon Entropy
(Rigau et al., 2008), and combined them to use as a
new aesthetic fitness score to evolve human
preferred images.
Likewise, authors (den Heijer and Eiben, 2011)
used two well-known aesthetic measures, Bell Curve
and GCF, as their fitness function to generate digital
art of vector graphics. In the same process, authors
(Bergen and Ross, 2012) used source image as an
aesthetic fitness measure by reading its colour pixels
to evolve an automatic vectorisation of that image.
In the evolution of art authors (den Heijer and
Eiben, 2010); (den Heijer and Eiben, 2011), used
four main various aesthetic measures: Benford’s
Law (Jolion, 2001), GCF, Information Theory and
Ross & Ralf’s Bell curve, to generate digital images
automatically. Many authors have applied aesthetic
measures to evolve various digital artefacts, 3D
structures (Bergen, 2011) (Bergen and Ross, 2012),
virtual creatures (Hornby and Pollack, 2001),
evolutionary art (Bergen and Ross, 2011), 3D art
(Pang and Hui, 2010), images (Romero et al., 2012)
etc. in recent decades.
Even though there is steady progress in applying
various aesthetic measures to evolve digital
artefacts, there is still much room for improvement
when compared with human evaluation. For
example, a survey was conducted by (Raffe et al.,
2012) on existing approaches of using evolutionary
algorithms for digital terrain generations that use
various fitness evaluations. Results showed both the
advantages and disadvantages of each approach.
Authors suggest that there is still need for robust
algorithms to evaluate aesthetics; in this case,
procedural terrain generation techniques, as none of
the existing tools can, at present, be practically used
for game development.
3 GCF AESTHETIC MEASURE
The main aesthetic measure used in this paper is
GCF, which we use to find the best computer
generated digital landscape designs by balancing the
contrast levels.
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
84
2.1 Global Contrast Factor
The main idea behind GCF (Matkovic et al., 2005) is
to take the average of various local contrast factors
and then compute a result with a weighting factor. It
starts by creating perceptual luminance L as shown
in Equation 1.
100∗
√
l
(1)
Where l is the linear luminance of each pixel with
applied gamma correction, = 2.2, as shown in
Equation 2.
l
k
255
(2)
Taking all 4 surrounding perceptual luminances, lc
i,
,
a local contrast factor C
is created using Equation 3.
lc
_
i|L
_
iL
_
i1||L
_
iL
_
i1| |L
_
i
L
_
iw| |L
_
iL
_
i
w|/4
(3)
C
i,
average local contrast factor is used to compute
the average of all the local contrast factors, lc
i,,
that
are produced at various resolutions. In our case, six
local contrast factors are performed.
C
1
w∗h
lc
(4)
At this stage the original image has been reduced to
half of its size in height and width by transforming
pixels to new ‘super pixels’. A super pixel is an
average of its surrounding pixel values. Figure 2
shows the size of the original image resolution
divided into 4 different resolutions when the
averaging of local contrast factors is performed.
Figure 2: Super Pixels stage at various resolutions.
In the final step, the GCF value is evaluated using
the summation of the entire average local contrast
factor, generated with weighting factors as shown in
Equation 5.
GCF w
∗C
(5)
Where,
w
0.406385∗
i
9
0.334573∗
i
9
0.0877526
(6)
The two images in Figure 3 show the GCF value
after performing the GCF aesthetic test.
Figure 3: GCF value of the left side image is 2.2695393
and right side image is 11.2654189.
4 PARAMETERS
The landscape designs used in our paper are
generated by use of a third party software
component called Terragen (Terragen, 2005). It
reads more than 800 parameter values in an xml
format called TGD and generates their graphical
representation as an image. We created a plugin
which generates an XML file with default and
altered parameter values which Terragen can read,
and thus generate landscapes accordingly. For
testing the evolution of landscape features, we
explore only 14 parameters in the evolutionary
process, based on the fitness evaluation.
Figure 4: Digitally generated scenery showing all the
parameters in details.
Table 1: Evolutionary parameter values range.
Parameter Min Max
Sun Elevation 0 90
Sun Heading 0 360
Terrain Height 2000 20000
Terrain Spikes 0 1
Cloud Altitude 5000 20000
Cloud Propagation Mix 0 1
Cloud Density 0 0.05
Cloud Depth 0 100
Water Waves 0 100
Water Roughness 0 0.3
Water Level -800 500
Sand Texture (RGB) 0 255
Rock Texture (RGB) 0 255
Grass Texture (RGB) 0 255
UseofGCFAestheticMeasureintheEvolutionofLandscapeDesigns
85
Parameters used in this paper include terrain
(height and spikiness), cloud (density, circulation,
depth and altitude), sun (elevation and heading),
water (roughness, wave height and level) and
textures (grass, sand and rock colours).
All parameter values are encoded using an 8-bit
binary representation for genetic operations and then
they are reverted back to their original values before
they are graphically represented. Floating point
values are converted to an 8-bit binary, index range
[0, 255] using Equation 7 and reverted to original
values using Equation 8.
Con
v
_
To
_
Bin
Input
Para
Ratio
(7)
_
_
∗
(8)
Where,
_
_
(9)
5 PROCESS
We implement an interactive tool that will evolve
landscape designs automatically based on an
aesthetic fitness score, GCF. An enhanced IGA is
also implemented to give the user direct access the
fitness measure used to evaluate landscape designs.
Users can overwrite the aesthetic measures with
their own ranking during the process.
Evolution of landscape design process is divided
into 5 core phases: Initial Population, Fitness
Evaluation, Selection, Genetic Operation and New
Population.
Figure 5: Evolutionary system – Flow chart.
5.1 Phase 1: Initialization
Our program generates a set of sixteen XML files
containing the randomised design parameters of
digital landscapes before they are rendered by the
scenery generator, Terragen. After rendering, sixteen
landscapes are presented by our GUI, as shown in
Figure 7.
5.2 Phase 2: Fitness Evaluation
The user is optionally allowed to rank their preferred
landscape designs to guide the evolution towards
their desired landscape. At this phase, the user
selects the best three parent templates from the GUI.
In an attempt to decrease user fatigue, we have given
the following options to help users evaluate the
landscapes:
5.2.1 Guidance during Evaluation
During user evaluation, each landscape is scored
using GCF, Kolmogorov aesthetic measure, or both
together, giving the top three ranked landscapes.
Users can follow these hints when in a dilemma over
which landscape to choose.
Note: When users select both the GCF and
Kolmogorov complexity aesthetic measures, the
mean value of both fitness scores is taken as the
final score and ranked accordingly using TotalScore
as shown in Equation 12.
_
∗
(10)
_
∗
(11)
⥤
_
2∗
(12)
5.2.1 Search via Computational Aesthetic
Measure
Users can pass control to the computational aesthetic
measures during the evolutionary process. At any
point in the process, users can set the number of
generations and allow the automated process of
computational aesthetic fitness measures to direct
the exploration of the fitness landscape. After those
set generations are evolved, the computational
aesthetic measure fitness evaluations are disengaged
and the control is handed back to the user for further
human evaluation.
5.3 Phase 3: Selection
The selection of parents is done using a ranking
based roulette wheel selection where the first
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
86
selected landscape, Rank 1, will have more
probability of being selected than the second, Rank
2, and more again than the third one, Rank 3. The
least probable selections are done on the remaining
landscape designs that have the least fitness values,
as shown in Figure 6.
Figure 6: Roulette wheel selection.
5.4 Phase 4: Genetic Operations
In this phase, a pair of selected parents produces two
new offspring with their own characteristics using
crossover and mutation operators.
5.4.1 Crossover
Each parameter value is converted to 8-bit binary
format before genetic operations take place. At a
random crossover point, random numbers of bits are
exchanged between each other to give two new
binary numbers offspring.
5.4.2 Mutation
In the mutation process, after the original value is
converted to binary format, at a random selection
point a binary bit is flipped over to its opposite
value, 1 to 0 or 0 to 1, giving random features.
5.5 Phase 5: New Population
A set of sixteen new XML files are generated,
defining the graphical properties of the new
offspring population. They are rendered by the
Terragen software component to produce novel
landscape designs and are then displayed in the GUI
for further evaluation. The whole process is repeated
until termination: when a user is satisfied or a set
number of generations are produced.
6 EXPERIMENTS AND RESUTLS
Two sets of experiments were performed.
The first experiment was conducted to test the
evolutionary search for digital landscapes using a
GCF aesthetic fitness function with default
parameter settings.
The second experiment was conducted to check the
efficiency of contrast levels over terrains only. This
was necessary to reduce the effect of cloud
reflection over water levels as seen in Figure 8.
6.1 Experiment 1: GCF Aesthetic Test
In the first experiment, the effectiveness of the GCF
fitness measure is tested. With a random distribution
of parameter values, within the range as shown in
Table 1, a set of 16 digital landscape designs are
generated with random colours in the initial phase as
shown in Figure 7. Then the GA is applied to
identify the top 3 digital images in each population;
to generate a new population automatically until the
required number of generations is met.
Figure 7: Initial set of randomly generated population of
Digital landscapes in search of GCF.
During the process of evolution, the GA selects the
parameters that maximizes the fitness factor and
adjusts them via crossover and mutation. In this case
the parameters such as cloud, water, terrain texture
and sun are adjusted to optimize the global contrast
factor in the generated digital images.
Even though the GCF fitness function reads and
adjusts the luminance of the pixels from a grey scale
image, the colour textures of the terrain are evolved
automatically by finding the right combination of
RGB values to match the contrast levels produced
by the GCF fitness function. The final generation of
this experiment is shown in Figure 8. Note that these
images contain high water levels as this tends to give
a high GCF score. The chart in Figure 9 shows that
there is an emerging trend of higher average fitness
scores over successive generations, which shows
UseofGCFAestheticMeasureintheEvolutionofLandscapeDesigns
87
that the evolutionary search mechanism is operating
as expected on the fitness function. Mutation
occasionally disrupts this upward trend to ensure an
element of diversity remains in the population.
Figure 8: Final generation of Digital landscapes in search
of GCF.
Figure 9: Experiment 1 – Fitness graph.
6.2 Experiment 2: GCF Aesthetic Test
with reduced Water and Clouds
Experiment 2 has the same goal as Experiment 1,
except in this case the effect of clouds and water are
reduced. This was done after feedback from experts
suggested that the reflective properties of water can
bias the contrast factor to high values.
Both the initial and final generation are shown in
Figure 10 and Figure 11 respectively. Again it can
be seen that the algorithm automatically directs the
evolution of landscapes towards a set of parameters
that represent balanced compositions. Interestingly,
the best-ranked generated landscapes tend to have
fairly naturalistic compositions compared to the
original generation, without any input from the user.
Table 2: Adjusted parameter values range
Parameter Min Max
Sun Elevation 0 90
Sun Heading 0 360
Terrain Height 2000 20000
Terrain Spikes 0 1
Cloud Altitude 5000 20000
Cloud Propagation Mix 0 1
Cloud Density 0 0.01
Cloud Depth 0 10
Water Waves 0 100
Water Roughness 0 0.3
Water Level -1200 -500
Sand Texture (RGB) Static Static
Rock Texture (RGB) Static Static
Grass Texture (RGB) Static Static
Figure 10: Initial generation of IGA test.
Figure 11: Final generation of IGA test.
5
6
7
8
9
1 3 5 7 9 1113151719212325
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
88
Figure 12: Experiment 2 – Fitness graph.
7 CONCLUSIONS AND FUTURE
WORK
Our objective in this study was to investigate GCF
aesthetic measures; to test GCF with our existing
tools, and verify the results as to whether or not they
can evolve towards balanced contrast levels. Both
aesthetic measures direct the evolutionary process
towards images that have balanced and naturalistic
characteristics. This would validate findings of other
researchers who employ similar aesthetic measures
for the auto-generation of art works and suggest that
there is some utility in computational fitness
evaluations. It is note-worthy that the images were
completely generated automatically and did not
require user input. This suggests that computational
aesthetic measures could be employed to reduce user
fatigue in interactive genetic algorithms, and perhaps
replace the user altogether. Further studies with
cohorts of real users will be planned to evaluate the
utility of this approach in future studies.
There is also an opportunity to add more fractal
terrain parameters into the GA process for achieving
more realistic digital artefacts. Moreover, adding
objects like plants, flowers, rocks and trees will give
a richer look to our output. We would also like to
implement more of our aesthetic measures and
investigate them with our existing ones in the
creation of evolutionary art.
REFERENCES
Alpine. (2011). Landscape Design using Interactive
Genetic Algorithm - target Alpine scenery . Available:
http://www.youtube.com/watch?v=OISauX-XAVY.
Last accessed 21st June 2013.
Assassin's Creed. (2007). Assassin's Creed IV Black
Flag. Available: http://assassinscreed.ubi.com/. Last
accessed 21st June 2013.
Bentley, P. (Ed.). (1999). Evolutionary design by
computers. Morgan Kaufmann. Bergen, S., & Ross, B.
J. (2012). Automatic and interactive evolution of
vector graphics images with genetic algorithms. The
Visual Computer, 28(1), 35-45.
Bergen, S. (2011). Automatic structure Generation using
genetic programming and fractal geometry. Brock
University, Ontario.
Bergen, S., & Ross, B. J. (2012). Aesthetic 3d model
evolution. In Evolutionary and Biologically Inspired
Music, Sound, Art and Design (pp. 11-22). Springer
Berlin Heidelberg.
Bergen, S., & Ross, B. J. (2011). Evolutionary art using
summed multi-objective ranks. In Genetic
Programming Theory and Practice VIII (pp. 227-244).
Springer New York.
Birkhoff, G. D. (1933). Aesthetic measure. Cambridge,
Mass.
Bryce. (2010). What is Bryce. Available:
http://www.daz3d.com/products/bryce/bryce-what-is-
bryce. Last accessed 21st June 2013.
Desert. (2011). Landscape Design using Interactive
Genetic Algorithm - target Desert scenery . Available:
http://www.youtube.com/watch?v=qDey8uyzvqg. Last
accessed 21st June 2013.
den Heijer, E., & Eiben, Á. E. (2011, July). Evolving art
with scalable vector graphics. In GECCO (pp. 427-
434).
den Heijer, E., & Eiben, A. E. (2010, July). Using
aesthetic measures to evolve art. In Evolutionary
Computation (CEC), 2010 IEEE Congress on (pp. 1-
8). IEEE.
den Heijer, E., & Eiben, A. E. (2011). Evolving art using
multiple aesthetic measures. In Applications of
Evolutionary Computation (pp. 234-243). Springer
Berlin Heidelberg.
Gimp. (1996). The GNU Image Manipulation
Program. Available: http://www.gimp.org/. Last
accessed 21st June 2013.
Halo Bungie. (2001). Halo Official Site. Available:
http://www.halowaypoint.com/en-us/. Last accessed
21st June 2013.
Hornby, G. S., & Pollack, J. B. (2001). Evolving L-
systems to generate virtual creatures. Computers &
Graphics, 25(6), 1041-1048.
Jolion, J. M. (2001). Images and Benford's law. Journal of
Mathematical Imaging and Vision, 14(1), 73-81.
Li, M. (1997). An introduction to Kolmogorov complexity
and its applications. Springer.
Li, Y., & Hu, C. J. (2010). Aesthetic learning in an
interactive evolutionary art system. In Applications of
Evolutionary Computation (pp. 301-310). Springer
Berlin Heidelberg.
Lightwave. (1993). Lightwave. Available:
https://www.lightwave3d.com/community/gallery/cate
gory/film/. Last accessed 21st June 2013.
Machado, P., & Cardoso, A. (1998). Computing
aesthetics. In Advances in Artificial Intelligence (pp.
219-228). Springer Berlin Heidelberg.
Matković, K., Neumann, L., Neumann, A., Psik, T., &
5
6
7
8
9
10
11
12
1 3 5 7 9 1113151719212325
UseofGCFAestheticMeasureintheEvolutionofLandscapeDesigns
89
Purgathofer, W. (2005, May). Global contrast factor-a
new approach to image contrast. In Proceedings of the
First Eurographics conference on Computational
Aesthetics in Graphics, Visualization and Imaging (pp.
159-167). Eurographics Association.
Pang, W., & Hui, K. C. (2010). Interactive evolutionary
3D fractal modeling. The Visual Computer, 26(12),
1467-1483.
Photoshop. (1990). Photo editor software. Available:
http://www.adobe.com/ie/products/photoshop.html.
Last accessed 21st June 2013.
Raffe, W. L., Zambetta, F., & Li, X. (2012, June). A
survey of procedural terrain generation techniques
using evolutionary algorithms. In Evolutionary
Computation (CEC), 2012 IEEE Congress on (pp. 1-
8). IEEE.
Rhythm and Hues Studio. (1987). Rhythm and Hues
Studio. Available: http://www.rhythm.com/home/.
Last accessed 21st June 2013.
Rigau, J., Feixas, M., & Sbert, M. (2008). Informational
aesthetics measures.Computer Graphics and
Applications, IEEE, 28(2), 24-34.
Romero, J., Machado, P., Carballal, A., & Correia, J.
(2012). Computing aesthetics with image judgement
systems. In Computers and Creativity (pp. 295-322).
Springer Berlin Heidelberg.
Ross, B. J., Ralph, W., & Zong, H. (2006). Evolutionary
image synthesis using a model of aesthetics.
In Evolutionary Computation, 2006. CEC 2006. IEEE
Congress on (pp. 1087-1094). IEEE.
Terragen. (2005). Terragen. Available: http://planetside.
co.uk/. Last accessed 21st June 2013.
Vue. (2012). e-on softwares. Available: http://www.e-
onsoftware.com/products/?page=%20artist. Last
accessed 21st June 2013.
Walsh, P., & Gade, P. (2010, July). Terrain generation
using an interactive genetic algorithm. In Evolutionary
Computation (CEC), 2010 IEEE Congress on(pp. 1-7).
IEEE.
Walsh, P., & Gade, P. (2011, June). The use of an
aesthetic measure for the evolution of fractal
landscapes. In Evolutionary Computation (CEC), 2011
IEEE Congress on (pp. 1613-1619). IEEE.
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
90
