Testing the Differences of using RGB and HSV Histograms during

Evolution in Evolutionary Art

P. Garc

ıa-S

anchez

, J. J. Merelo

, D. Calandria

, A. B. Pelegrina

, R. Morcillo

, F. Palacio

and R. H. Garc

ıa-Ortega

Dept. of Computer Architecture and Computer Technology and CITIC-UGR, ETS. Inform

atica y Telecomunicaci

on,

University of Granada, Granada, Spain

Proemium, Granada, Spain

Department of Translation and Interpreting, University of Granada, Granada, Spain

ETS. Inform

atica y Telecomunicaci

on, University of Granada, Granada, Spain

Fundaci

on I+D del Software Libre, Granada, Spain

Keywords:

Evolutionary Art, HSV, RGB, Processing.

Abstract:

This paper compares the use of RGB and HSV histograms during the execution of an Evolutionary Algorithm.

This algorithm generates abstract images that try to match the histograms of a target image. Three different

ﬁtness functions have been used to compare: the differences between the individual with the RGB histogram

of the test image, the HSV histogram, and an average of the two histograms at the same time. Results show

that the HSV ﬁtness also increases the similarities of the RGB (and therefore, the average) more than the other

two measures.

1 INTRODUCTION

Evolutionary Art (Corne and Bentley, 2001) is a

branch of generative art (Fernandes et al., 2012) cre-

ated using a computer, following the principle of the

survival of the ﬁttest, using Evolutionary Computa-

tion methods (Eiben and Smith, 2005). A population

of artistic works (individuals) are evaluated with an

aesthetic measure to yield a score (ﬁtness). These in-

dividuals are combined and mutated to generate an

offspring with inherited properties of the parents, dur-

ing a certain number of times.

There exist several metrics to score the gener-

ated images, such as opinion from humans, or image

characteristics (for example, speciﬁc combination of

colours). The main goal of this paper is to study the

differences of using the information of the HSV (Hue,

Saturation, Value) and RGB (Red, Green, Blue) his-

tograms during the evolution. Although the two his-

tograms represent the same information, using HSV

(instead as RGB) as a color model increases the ac-

curacy in image retrieval and indexing, as explained

in (Sebe and Lew, 2000). The results of this investi-

gation can help in future evolutionary art algorithms,

adding the most appropiate color model feature with

other features available in the literature. For example,

to be used as one of the features of any kind of classi-

ﬁer. In this work the Processing (Reas and Fry, 2007)

framework is used inside an Evolutionary Algorithm

(EA) to model the individuals, generate their associate

images and extract information of them (HSV, RGB

and Average histograms) to ﬁt with the histograms of

a test image. Processing has been integrated in the

OSGiLiath (Garc

ıa-S

anchez et al., 2013) evolution-

ary framework to take advantage of the capabilities

offered in image manipulation and analysis.

The rest of the work is organized as follows: in

Section 2, a brief review on Evolutionary Art is pre-

sented. Processing framework and image informa-

tion are described next (Section 3). The experimental

setup and results are presented in Sections 4 and 5,

respectively. Finally, the conclusions and future work

can be found in Section 6.

2 STATE OF THE ART

Computational Aesthetics “is the research of compu-

tational methods that can make applicable aesthet-

168

García-Sánchez P., Merelo J., Calandria D., Pelegrina A., Morcillo R., Palacio F. and García-Ortega R..

Testing the Differences of using RGB and HSV Histograms during Evolution in Evolutionary Art.

DOI: 10.5220/0004629701680174

In Proceedings of the 5th International Joint Conference on Computational Intelligence (ECTA-2013), pages 168-174

ISBN: 978-989-8565-77-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

ics decisions in a similar fashion as humans can”

(Hoenig, 2005). In the ﬁeld of computational aesthet-

ics, evolutionary systems can play an important role,

by enabling the evolution of aesthetically pleasing

or innovative structures (DiPaola and Gabora, 2009).

Evolutionary art is characterized by the use of evolu-

tionary principles and natural selection as a genera-

tive process. One of the earliest applications of evo-

lutionary systems to generate art is the proposal of

Sims to use an EA to create complex images (Sims,

1991) or virtual creatures (Sims, 1994). In evolution-

ary art systems, the evaluation of the aesthetics can

be done using human feedback, with some interactive

evaluation of the population, such as (Ashlock, 2006;

Draves, 2006; Moroni et al., 2000) and (Sims, 1991).

It also can be achieved by using an automatic evalu-

ation of ﬁtness, as presented in (Aguilar and Lipson,

2008; Den Heijer and Eiben, 2010; DiPaola and Gab-

ora, 2009; Li et al., 2012), and (Sims, 1994).

One of the main challenges in Evolutionary Art is

how to measure aesthetic value of a piece of evolu-

tive art. The source of this difﬁculty lies in the in-

herent complexity, subjectivity and dynamism of aes-

thetics. Nevertheless, a wide number of metrics has

been presented. According to (Galanter, 2012), these

measures can be classiﬁed into several categories in

several pieces of research. The ﬁrst category involves

the evaluation of the aesthetics of a piece of art by a

formula or principle (e.g., pythagorean proportions).

Other measures apply certain principles of design,

such as the rule of thirds or color theory (e.g., us-

ing complimentary colors in Pop Art (den Heijer and

Eiben, 2012)), neural networks or complexity mea-

sures.

This classiﬁcation also provides a sub-

classiﬁcation for evolutionary systems. First, it

identiﬁes interactive evolutionary systems, where the

ﬁtness of the individuals is determined by human

agents. Another category is performance based goals:

certain properties of the art piece are evaluated and

optimized based on performance measures (e.g.,

usable surface in a furniture design generator). Other

systems use an exemplar (i.e., real world example)

as a way to measure the ﬁtness of the individuals.

Finally, some models use the idea that the complexity

is directly related to aesthetics and follow the path

ﬁrstly stablished in (Birkhoff, 2003). Given the

multidimensional nature of aesthetics judgement,

multi-objective EAs are a straightforward option to

deal with this multidimensionality. Other extensions

to EA, like coevolution or agent swarm behavior, can

be used in evolutionary art systems.

A brief classiﬁcation of the aesthetic measures

found in the evolutionary art systems mentioned in

the previous paragraph is shown in Table 1.

Several methods for the representation of the art

in evolutionary art have been proposed. In symbolic

expression, the genotype is a tree of expressions and

the phenotype consists in the image produced by the

evaluation of the tree. Shape grammars can also be

used as a formal description of the image. Previously

existing images can be used as a basis for the evo-

lution process. Finally, other representations can be

based on mathematical models, like fractals or cellu-

lar automata.

3 PROCESSING AND

HISTOGRAMS

In this section we describe Processing

and the his-

tograms used. Processing (Reas and Fry, 2007) is

a framework formed by a simple programming lan-

guage and an integrated development environment

(IDE) primarily created for electronic and visual

artists, designers, musicians, etc. Processing offers

the following advantages:

• Processing was created for artists, rather than pro-

grammers. So, it allows very complex draw-

ings and interactive applications with few lines of

code.

• It is an Open Source software (licensed under the

GNU Lesser General Public License), and counts

with a large development community.

• It is based on OpenGL, thus providing 3D accel-

eration.

• It also includes more than 100 libraries for video,

sound, physics, computer vision, networking, etc.

• Easy integration with Java, HTML5 and Android.

However, being a light framework, there exist

some disadvantages:

• More complex applications require more pro-

gramming skills.

• The calculations of large computer images are a

bit inefﬁcient (although expert programmers can

manage OpenGL at low level to ﬁx this).

There exist a lot of interactive artistic projects

made with Processing; examples include art

generation, artiﬁcial life, interactive music

and other. A good selection can be seen in

http://processing.org/exhibition/.

The Color module can be used to analyze images

taking into account their histogram. The color his-

togram represents the frequency of occurrence of each

http://www.processing.org/

TestingtheDifferencesofusingRGBandHSVHistogramsduringEvolutioninEvolutionaryArt

169

Table 1: Classiﬁcation of the aesthetic measures used in a brief review of the literature on evolutive art.

Type Aesthetic Measure

Formulaic and Geometric Theories Fractal dimension (Den Heijer and Eiben, 2010), Image order (Li

et al., 2012), Benford Law (Del Acebo and Sbert, 2005)

Based on Design Principles Color contrast (hue) (den Heijer and Eiben, 2012), Color ingredi-

ent (Li et al., 2012), Composition, tonality and color (DiPaola and

Gabora, 2009).

Interactive Evolutionary Computation The electric sheep project (Draves, 2006) (Ashlock, 2006; Moroni

et al., 2000)

Error relative to Exemplars Resemblance score (DiPaola and Gabora, 2009), pixel comparation

(Aguilar and Lipson, 2008)

Performance based goals Evolving virtual creatures (Sims, 1994)

Complexity measures Image complexity (Li et al., 2012), Machado and Cardoso aesthetic

measure (Machado and Cardoso, 1998)

color intensities present in the image, by accounting

for such sharing pixels color intensity values.

The histogram is composed of different ranges or

bins that represent a value or set of values of color

intensity. The color space is deﬁned as a model

representation with respect to color intensity values.

Two color models are used in this paper: RGB (Red,

Green, Blue) and HSV (Hue, Saturation and Value).

The RGB model is an additive color model in which

red, green and blue are added together in different

proportions to reproduce a wide range of colors, while

the HSV is based on hue or tone, saturation and

brightness. While the RGB model is the closest to the

way color is processed in some machines, the HSV

representation provides a more accurate way to model

how humans perceive colors, and also provides more

information in image retrieval (Sebe and Lew, 2000).

Figure 1 shows the RGB histogram of the image in

Figure 2 (photo taken by the ﬁrst author).

4 EXPERIMENTAL SETUP

This section shows how Processing has been used in

the EA, the individual representation, the ﬁtness func-

tions, and the parameters of the experiments.

4.1 Integrating Processing in Java

Processing can be integrated with Java just by adding

a jar (a Java library) to existing software. In this

work, Processing has been integrated to an existent

EA framework, OSGiLiath (Garc

ıa-S

anchez et al.,

2013), a service-oriented framework based on Java

that includes a lot of primitives and services for Evo-

lutionary Computation. A new module called OS-

Figure 1: RGB histogram of Figure 2.

Figure 2: Test image to compare with the Fitness functions

of our algorithm.

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

170

GiLiART has been added to the publicly-available

source code of OSGiLiath (available in

http://www.

osgiliath.org

) under a LGPL License. Then, us-

ing the packages available in the Processing library

the EA can generate individuals, manipulate images

or extract information.

4.2 Individual Representation

To perform the experiments, the genome of the indi-

vidual is a list of circles. Each circle has a position,

radium and color. This list can be recombined or mu-

tated (changing the color, position or radium of a cir-

cle of the list).

4.3 Fitness used

For this piece of research, we focused on the aes-

thetics measure of histogram comparison. The ﬁtness

functions are included in the “Error relative to Exem-

plars” category, using Galanter (Galanter, 2012) clas-

siﬁcation. The idea is to obtain an image with the

same proportion of tones and colors of a aesthetical

existent image.

Three different ﬁtness functions have been tested:

• RGB Difference: The difference of the RGB his-

togram of the individual with the RGB histogram

of the test image.

• HSV Difference: The difference of the HSV his-

togram of the individual with the HSV histogram

of the test image.

• Average Difference: An average of the two previ-

ous differences.

The range of the these ﬁtness function has been

normalized to vary from 0 (totally different his-

tograms) to 1 (the same histogram).

For every color property (i.e., RED, GREEN,

BLUE, HUE, SATURATION and VALUE), the his-

togram is computed using the expression (1) for each

possible value (0-255). Then, again for every prop-

erty, the difference between the target image and the

individual histograms is obtained using (2). Finally,

the three ﬁtness are calculated: RGB ﬁtness (6), HSV

ﬁtness (10) and AVERAGE ﬁtness (11).

H(c, prop) =

∑

j=0

{

1 prop( j ) = c

0 otherwise

(1)

di f f (h

, h

) =

255

∑

j=0

( j) − h

( j)| (2)

(i) = di f f (H(i, RED), H(target, RED)) (3)

(i) = di f f (H(i, GREEN), H(target, GREEN)) (4)

(i) = di f f (H(i, BLUE), H(target, BLUE)) (5)

f itness

RGB

(i) = 1 − 128

(i) + d

(i)

(6)

(i) = di f f (H(i, HUE), H(target, HUE)) (7)

(i) = di f f (H(i, SAT ), H(target, SAT )) (8)

(i) = di f f (H(i, VAL), H(target, VAL)) (9)

f itness

HSV

(i) = 1 − 128

(i) + d

(i)

(10)

f itness

AV ERAGE

(i) =

f itness

RGB

+ f itness

HSV

(11)

4.4 Parameters used

A steady-state evolutionary algorithm has been used.

Each individual is randomly generated at the initial-

ization of the EA. The genome size is 50 elements

(circles of maximum radium of 128 pixels). Popu-

lation size has been set to 32 individuals. Uniform

crossover rate is 0.5, and a binary tournament has

been chosen for selection (that is, a pool of 16 par-

ents is selected and crossed). Mutation probability is

0.04 (the usual value of 1/genomesize). Finally, the

image size for each individual is 256x256 pixels. The

individuals have been compared with the histograms

obtained from the image of Figure 2 to guide the evo-

lution.

5 RESULTS

Because of the stochastic nature of the EAs, each al-

gorithm has been executed 30 times for each differ-

ent ﬁtness. Table 2 shows the average differences

(and standard deviation) attained with each ﬁtness

used. As can be seen, using the HSV histogram dif-

ferences as ﬁtness produces a higher RGB similarity

(and therefore, average) than using the RGB or Av-

erage ﬁtness. However, using the average between

the two histogram differences produces higher sim-

ilarity in HSV (0.294) than only taking into account

the HSV. The maximum ﬁtness is around 25% of sim-

ilarity with the original image since the individual is

a list of 50 circles, and therefore, only a maximum

of 50 different colors are used (while in the origi-

nal jpg image can be more than millions). See the

histogram of a generated best individual by the algo-

rithm in Figure 3. An example of evolution for each

ﬁtness can be seen in Figure 4, 5 and 6. Comparing

TestingtheDifferencesofusingRGBandHSVHistogramsduringEvolutioninEvolutionaryArt

171

Figure 3: RGB histogram of a solution generated by the

algorithm.

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0 10 20 30 40 50 60 70 80

Fitness

Generations

Using RGB Histogram Fitness

RGB

HSV

AVERAGE

Figure 4: Evolution of the difference in RGB histogram of

the best individual compared with the test image.

with the RGB histogram as ﬁtness, a bigger ﬂuctua-

tion in the HSV is produced (Figure 4). This can be

explained because the RGB information tends to be

more noisy than HSV information: in fact, in (Sebe

and Lew, 2000) authors explain the problems this his-

togram offers with respect to HSV in image retrieval.

Although there is the same information modeled in

both histograms, the transformation from one to an-

other is not linear, so there is no relation with the his-

tograms of individuals generated during the evolution.

The best individuals attained are shown in Figure

7. Note that, although the numeric ﬁtness is similar,

they produce different color tones. This can be ex-

plained for the limitation of colors used in the indi-

vidual representation, as previously said, or the noisy

characteristic of the RGB histogram. Figure 8 shows

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

0 10 20 30 40 50 60 70 80

Fitness

Generations

Using HSV Histogram Fitness

RGB

HSV

AVERAGE

Figure 5: Evolution of the difference in HSV histogram of

the best individual compared with the test image.

0.1

0.15

0.2

0.25

0.3

0.35

0 10 20 30 40 50 60 70 80

Fitness

Generations

Using Average Histogram Fitness

RGB

HSV

AVERAGE

Figure 6: Evolution of the difference of average of RGB

and HSV histogram of the best individual compared with

the test image.

one evolution of the best individual using the HSV

ﬁtness in the ﬁrst 64 generations.

6 CONCLUSIONS AND FUTURE

WORK

This paper introduces an Evolutionary Algorithm that

uses the Processing framework to generate images

and to extract image information using HSV and RGB

histograms. In this work individuals are represented

as a list of Processing primitives (circles) and the ﬁt-

ness functions used are based on the similarity with an

existent aesthetic image. Three different ﬁtness func-

tions using color histogram have been tested: differ-

ence between the HSV and RGB histograms, and an

average difference of the two histograms at the same

time. Experiments show that better results in terms

of similarity are obtained using the HSV comparison

(due to the noisy information provided by the RGB).

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

172

Table 2: Results for the different ﬁtness (average of the 30 executions and standard deviation). Only one histogram type is

used for ﬁtness calculation, but the other values obtained are also added.

Differences used in Fitness Obtained RGB Obtained HSV Obtained Average

RGB Histogram 0.267 ± 0.012 0.170 ± 0.010 0.218 ± 0.009

HSV Histogram 0.227 ± 0.017 0.265 ± 0.021 0.246 ± 0.010

Average Histogram 0.173 ± 0.012 0.294 ± 0.013 0.234 ± 0.010

(a) Best individual using RGB. (b) Best individual using HSV. (c) Best individual using AVERAGE.

Figure 7: Best individuals obtained with the three ﬁtness used (HSV, RGB and AVERAGE).

Figure 8: Evolution of the best individual using the HSV

histogram difference.

The future work for this research also includes

more experiment with other kind of individuals,

apart from circles: using other primitives, such as

rectangles or triangles, for example. The use of tex-

tures and gradients will generate images with higher

number of colors, obtaining more ﬁdelity (more than

25%) with the test image. Other metrics explained in

previous sections will be also implemented. Finally,

our intention is not only to create only static images,

but use the Processing libraries to create evolution-

ary interactive art combining sounds and motion.

A human guidance tool is also being developed

to obtain human feedback to create a knowl-

edge base for future experimentation (available in

http://evorq.ugr.es:8080/HumanGuidance).

The results will be gathered to create a database

of different features to guide the evolution. More

complex measurements will be studied in next works,

taking into account that the HSV is the color mode

that provides more information during the evolution,

having less noisy behaviour.

The used software and algorithms presented are

Open Source under a GPL license, and can be ob-

tained from

http://www.osgiliath.org.

ACKNOWLEDGEMENTS

This work has been supported in part by FPU

research grant AP2009-2942 and projects EvOrq

(TIC-3903), SINECA (0100DGT21285, Spanish Di-

reccion General de Traﬁco), TIN2011-28627-C04-

02 and FFI2011-22397 (funded by Spanish Min-

istry for Economy and Productivity). This work

was developed during the 5 Hackathon of the

Ofﬁce of Open Software of the University of

Granada

http://sl.ugr.es/5hackathon,

where

several members of different disciplines collaborated

during its creation.

TestingtheDifferencesofusingRGBandHSVHistogramsduringEvolutioninEvolutionaryArt

173

REFERENCES

Aguilar, C. and Lipson, H. (2008). A robotic system for

interpreting images into painted artwork. In Proceed-

ings of the 11th Generative Art Conference (GA2008),

Politecnico di Milano University, Milan, Italy.

Ashlock, D. (2006). Evolutionary exploration of the man-

delbrot set. In Evolutionary Computation, 2006. CEC

2006. IEEE Congress on, pages 2079–2086. IEEE.

Birkhoff, G. (2003). Aesthetic Measure 1933. Harvard Uni-

versity Press.

Corne, D. W. and Bentley, P. J. (2001). Creative evolution-

ary systems. Morgan Kaufmann.

Del Acebo, E. and Sbert, M. (2005). Benford’s law for natu-

ral and synthetic images. In Computational Aesthetics

in Graphics, Visualization and Imaging, pages 169–

176. The Eurographics Association.

Den Heijer, E. and Eiben, A. (2010). Comparing aesthetic

measures for evolutionary art. Applications of Evolu-

tionary Computation, pages 311–320.

den Heijer, E. and Eiben, A. (2012). Evolving pop art using

scalable vector graphics. Evolutionary and Biologi-

cally Inspired Music, Sound, Art and Design, pages

48–59.

DiPaola, S. and Gabora, L. (2009). Incorporating charac-

teristics of human creativity into an evolutionary art

algorithm. Genetic Programming and Evolvable Ma-

chines, 10(2):97–110.

Draves, S. (2006). The electric sheep. ACM SIGEVOlution,

1(2):10–16.

Eiben, A. and Smith, J. (2005). What is an evolutionary

algorithm? In Rozenberg, G., editor, Introduction to

Evolutionary Computing, pages 15–35. Addison Wes-

ley.

Fernandes, C. M., Garc

ıa, A. M., Guerv

os, J. J. M., and

Rosa, A. C. (2012). Pherogenic drawings - generat-

ing colored 2-dimensional abstract representations of

sleep eeg with the kants algorithm. In IJCCI 2012 -

Proceedings of the 4th International Joint Conference

on Computational Intelligence, Barcelona, Spain, 5 -

7 October, 2012, pages 72–80.

Galanter, P. (2012). Computational aesthetic evaluation:

past and future. In Computers and Creativity , pages

255–293. Springer.

Garc

ıa-S

anchez, P., Gonz

alez, J., Castillo, P., Arenas, M.,

and Merelo-Guerv

os, J. (2013). Service oriented evo-

lutionary algorithms. Soft Computing, 17(6):1059–

1075.

Hoenig, F. (2005). Deﬁning computational aesthetics. In

Proceedings of the First Eurographics conference on

Computational Aesthetics in Graphics, Visualization

and Imaging, Computational Aesthetics’05, pages

13–18, Aire-la-Ville, Switzerland, Switzerland. Euro-

graphics Association.

Li, Y., Hu, C., Chen, M., and Hu, J. (2012). Investigating

aesthetic features to model human preference in evo-

lutionary art. Evolutionary and Biologically Inspired

Music, Sound, Art and Design, pages 153–164.

Machado, P. and Cardoso, A. (1998). Computing aesthetics.

Advances in Artiﬁcial Intelligence, pages 105–119.

Moroni, A., Manzolli, J., Zuben, F. V., and Gudwin, R.

(2000). Vox populi: An interactive evolutionary sys-

tem for algorithmic music composition. Leonardo

Music Journal, pages 49–54.

Reas, C. and Fry, B. (2007). Processing: A Programming

Handbook for Visual Designers and Artists. Working

paper series (National Bureau of Economic Research).

Mit Press.

Sebe, N. and Lew, M. S. (2000). A maximum likelihood

investigation into color indexing. Proceedings Visual

Interface, pages 101–106.

Sims, K. (1991). Artiﬁcial evolution for computer graphics.

Computer Graphics, 25(4).

Sims, K. (1994). Evolving virtual creatures. In Proceedings

of the 21st annual conference on Computer graphics

and interactive techniques, pages 15–22. ACM.

IJCCI2013-InternationalJointConferenceonComputationalIntelligence

174