Evolving Figurative Images Using Expression-Based Evolutionary Art
Joao Correia ˜ and Penousal Machado
CISUC, Department of Informatics Engineering
University of Coimbra
3030 Coimbra, Portugal
jncor@dei.uc.pt, machado@dei.uc.pt
Juan Romero and Adrian Carballal
Faculty of Computer Science
University of A Coruna˜
Coruna, Spain ˜
jj@udc.es, adrian.carballal@udc.es
Abstract
The combination of a classifier system with an evolutionary
image generation engine is explored. The framework is composed
of an object detector and a general purpose, expressionbased,
genetic programming engine. Several object detectors
are instantiated to detect faces, lips, breasts and leaves.
The experimental results show the ability of the system to
evolve images that are classified as the corresponding objects.
A subjective analysis also reveals the unexpected nature and
artistic potential of the evolved images.
Introduction
Expression based Evolutionary Art (EA) systems have, in
theory, the potential to generate any image (Machado and
Cardoso 2002; McCormack 2007). In practice, the evolved
images depend on the representation scheme used. As a consequence,
the results of expression-based EA systems tend
to be abstract images. Although this does not represent a
problem, there is a desire to evolve figurative images by evolutionary
means since the start of EA. An early example of
such an attempt can be found in the work of Steven Rooke
(World 1996).
McCormack (2005; 2007) identified the problem of finding
a symbolic-expression that corresponds to a known “target”
image as one of the open problems of EA. More exactly,
the issue is not finding a symbolic-expression, since
this can be done trivially as demonstrated by Machado and
Cardoso (2002), the issue is finding a compact expression
that provides a good approximation of the “target” image
and that takes advantage of its structure. We address this
open problem by generalizing the problem – i.e., instead of
trying to match a target image we evolve individuals that
match a given class of images (e.g. lips).
The issue of evolving figurative images has been tackled
by two main types of approach: (i) Developing tailored EA
systems which resort to representations that promote the discovery
of figurative images, usually of a certain kind; (ii)
Using general purpose EA systems and developing fitness
assignment schemes that guide the system towards figurative
images. In the scope of this paper we are interested in
the second approach.
Romero et al. (2003) suggest combining a general purpose
evolutionary art system with an image classifier trained
to recognize faces, or other types of objects, to evolve
images of human faces. Machado, Correia, and Romero
(2012a) presented a system that allowed the evolution of
images resembling human faces by combining a generalpurpose,
expression-based, EA system with an off-the-shelf
face detector. The results showed that it was possible
to guide evolution and evolve images evocative of human
faces.
Here, we demonstrate that other classes of object can
be evolved, generalizing previous results. The autonomous
evolution of figurative images using a general purpose EC
system has rarely been accomplished. As far as we know,
evolving different types of figurative images using the same
expression-based EC system and the same approach has
never been accomplished so far (with the exception of userguided
systems).
We show that this can be attained with off-the-shelf classifiers
classifiers, which indicates that the approach is generalizable,
and also with purpose-built ones, which indicates
that it is relatively straightforward to customize it to specific
needs. We chose a rather ad-hoc set of classifiers in an attempt
to demonstrate the generality of the approach.
The remainder of this paper is structured as follows: A
brief overview of the related work is made in the next section;
Afterwards we describe the approach for the evolution
of objects describing the framework, the Genetic Programming
(GP) engine, the object detection system, and fitness
assignment; Next we explain the experimental setup, the results
attained and their analysis and; Finally we draw overall
conclusions and indicate future research.
Related Work
The use of Evolutionary Computation (EC) for the evolution
of figurative images is not new. Baker (1993) focuses on the
evolution of line drawings, using a GP approach. Johnston
and Caldwell (1997) use a Genetic Algorithm (GA) to recombine
portions of existing face images, in an attempt to
build a criminal sketch artist. With similar goals, Frowd,
Hancock, and Carson (2004) use a GA, Principal Components
Analysis and eigenfaces to evolve human faces. The
evolution of cartoon faces (Nishio et al. 1997) and cartoon
face animations (Lewis 2007) through GAs has also been explored.
Additionally, Lewis (2007) evolved human figures.
The previously mentioned approaches share two common
aspects: the systems have been specifically designed for the
Proceedings of the Fourth International Conference on Computational Creativity 2013 24
evolution a specific type of image; the user guides evolution
by assigning fitness. The work of Baker (1993) is an
exception, the system can evolve other types of line drawings,
however it is initialized with hand-built line drawings
of human faces.
These approaches contrast with the ones where general
purpose evolutionary art tools, which have not been designed
for a particular type of imagery, are used to evolve
figurative images. Although the images created by their
systems are predominantly abstract, Steven Rooke (World
1996) and Machado and Romero (see, e.g., 2011), among
others, have successfully evolved figurative images using
expression-based GP systems and user guided evolution.
More recently, Secretan et al. (2011) created picbreeder, a
user-guided collaborative evolutionary engine. Some of the
images evolved by the users are figurative, resembling objects
such as cars, butterflies and flowers.
The evolution of figurative images using hardwired fitness
functions has also been attempted. The works of by
Ventrella (2010) and DiPaola and Gabora (2009) are akin
to a classical symbolic regression problem in the sense that
a target image exists and the similarity between the evolved
images and the target image is used to assign fitness. In addition
to similarity, DiPaola and Gabora (2009) also consider
expressiveness when assigning fitness. This approach results
in images with artistic potential, which was the primary goal
of these approaches, but that would hardly be classified as
human faces. As far as we know, the difficulty to evolve a
specific target image, using symbolic regression inspired approaches,
is common to all “classical” expression-based GP
systems.
The concept of using a classifier system to assign fitness
is also a researched topic: in the seminal work of Baluja,
Pomerlau, and Todd (1994) an Artificial Neural Network
trained to replicate the aesthetic assessments is used; Saunders
and Gero (2001) employ a Kohonen Self-Organizing
network to determine novelty; Machado, Romero, and Manaris
(2007) use a bootstrapping approach, relying on a neural
network, to promote style changes among evolutionary runs;
Norton, Darrell, and Ventura (2010) train Artificial Neural
Networks to learn to associate low-level image features to
synsets that function as image descriptors and use the networks
to assign fitness.
Overview of the Approach
Figure 1 depicts an overview of the framework, which is
composed of two main modules, an evolutionary engine and
a classifier.
The approach can be summarized as follows:
1. Random initialization of the population;
2. Rendering of the individuals, i.e., genotype-phenotype
mapping;
3. Apply the classifier to each phenotype;
4. Use the results of the classification to assign fitness; This
may require assessing internal values and intermediate results
of the classification;
Figure 1: Overview of the system.
5. Select progenitors; Apply genetic operators, create descendants;
Use the replacement operator to update the
current population;
6. Repeat from 2 until some stopping criterion is met.
The framework was instantiated with a general-purpose
GP-based image generation engine and with a Haar Cascade
Classifier. To create a fitness function able to guide evolution
it is necessary to convert the binary output of the detector
to one that can provide suitable fitness landscape. This
is attained by accessing internal results of the classification
task that give an indication of the degree of certainty in the
classification. In the following sections we explain the components
of the framework, namely, the evolutionary engine,
the classifier and the fitness function.
Genetic Programming Engine
The EC engine used in these experiments is inspired by the
works of Sims (1991). It is a general purpose, expressionbased,
GP image generation engine that allows the evolution
of populations of images. The genotypes are trees composed
of a lexicon of functions and terminals. The function set is
composed of simple functions such as arithmetic, trigonometric
and logic operations. The terminal set is composed
of two variables, x and y, and randomly initialized constants.
The phenotypes are images that are rendered by evaluating
the expression-trees for different values of x and y, which
serve both as terminal values and image coordinates. In
other words, to determine the value of the pixel in the (0,0)
coordinates one assigns zero to x and y and evaluates the
expression-tree (see figure 2). A thorough description of the
GP engine can be found in (Machado and Cardoso 2002).
Figure 3 displays typical imagery produced via interactive
evolution using this EC system.
Object Detection
For classification purposes we use Haar Cascade classifiers
(Viola and Jones 2001). The classifier assumes the form of a
cascade of small and simple classifiers that use a set of Haar
features (Papageorgiou, Oren, and Poggio 1998) in combination
with a variant of the Adaboost (Freund and Schapire
1995), and is able to attain efficient classifiers. This classi-
fication approach was chosen due to its state of the art relevance
and for its fast classification. Both code and executables
are integrated in the OpenCV API1.
The face detection process can be summarized as follows:
1
OpenCV — http://opencv.org/
Proceedings of the Fourth International Conference on Computational Creativity 2013 25
Figure 2: Representation scheme with examples of functions
and the corresponding images.
Figure 3: Examples of images generated by the evolutionary
engine using interactive evolution.
1. Define a window of size w (e.g. 20 ⇥ 20).
2. Define a scale factor s greater than 1. For instance 1.2
means that the window will be enlarged by 20%.
3. Define W and H has the size of the input image.
4. From (0, 0) to (W, H) define a sub-window with a starting
size of w for calculation.
5. For each sub-window apply the cascade classifier. The
cascade has a group of stage classifiers, as represented in
figure 4. Each stage is composed, at its lower level, of
a group of Haar features 5. Apply each feature of each
corresponding stage to the sub-window. If the resulting
value is lower than the stage threshold the sub-window is
classified as a non-object and the search terminates for the
sub-window. If it is higher continue to next stage. If all
cascade stages are passed, the sub-window is classified as
!"#$%&'()*
+,-."/,0
1
+,-."/,0
2
3
+,-."/,0
1
+,-."/,0
2
3
+
4#5,6.07,.,6.,(
!.-8,01 !.-8,02
+
20!.-8,9
2)'04#5,6.
Figure 4: Cascade of classifiers with N stages, adapted from
(Viola and Jones 2001).
Figure 5: The set of possible features, adapted from (Lienhart
and Maydt 2002).
containing an object.
6. Apply the scale factor s to the window size w and repeat
5 until window size exceeds the image in at least one dimension.
Fitness Assignment
The process of fitness assignment is crucial from an evolutionary
point of view, and therefore it holds a large importance
for the success of the described system. The goal is to
evolve images that the object detector classifies as an object
of the positive class. However, the binary output of the detector
is inappropriate to guide evolution. A binary function
gives no information of how close an individual is to being
a valid solution to the problem and, as such, the EA would
be performing, essentially, a random search. It is necessary
to extract additional information from the classification detection
process in order to build a suitable fitness function.
This is attained by accessing internal results of the classi-
fication task that give an indication of the degree of certainty
in the classification. Based on results of past experiments
(Machado, Correia, and Romero 2012a; 2012b) we employ
the following fitness function:
f itness(x) =
nstages
X x
i
stagedifx(i)⇤i+nstagesx⇤10 (1)
The underlying rational is the following: images that
go through several classification stages, and closer to be
classified as an object, have higher fitness than those rejected
in early stages. Variables nstagesx and stagedifx(i)
Proceedings of the Fourth International Conference on Computational Creativity 2013 26
Table 1: Haar Training parameters.
Parameter Setting
Number of stages 30
Min True Positive rate per stage 99.9%
Max False Positive rate per stage 50%
Object Width 20 or 40(breasts,leaf)
Object Height 20 or 40(leaf)
Haar Features ALL
Number of splits 1
Adaboost Algorithm GentleAdaboost
are extracted from the object detection algorithm. Variable
nstagesx, holds the number of stages that image, x, has successfully
passed. That is, an image that passes several stages
is likely to be closer of being recognized as having a object
than one that passes fewer stages. In other words, passing
several stages is a pre-condition to be classified as having
the object. Variable stagedifx(i) holds the maximum difference
between the threshold necessary to overcome stage i
and the value attained by the image at the i
th stage. Images
that are clearly above the thresholds are preferred over ones
that are only slightly above them. Obviously, this fitness
function is only one of the several possible ones.
Experimentation
Within the scope of this paper we intend to evolve the following
objects: faces, lips, breasts and leaves. For the first
two we use off-the-shelf classifiers that were already trained
and used by other researchers in different lines of investigation
(Lienhart and Maydt 2002; Lienhart, Kuranov, and Pisarevsky
2003; Santana et al. 2008). For the last two we created
our own classifiers, by choosing suitable datasets and
training the respective object classifier.
In order to construct an object classifier we need to construct
two datasets: (i) positive – examples of images that
contain the object we want to detect; (ii) negative – images
that do not contain the object. Furthermore, for the positive
examples, we must identify the location of the object in the
images (see figure 6) in order to build the ground truth file
that will be used for training.
For these experiments, the negative dataset was attained
by picking images from a random search using image search
engines, and from the Caltech-256 Object Category dataset
(Griffin, Holub, and Perona 2007). Figure 7 depicts some
of the images used as negative instances. In what concerns
the positive datasets: the breast object detector was built by
searching images on the web; the leaf dataset was obtained
from the Caltech-256 Object Category dataset and from web
searches. As previously mentioned, the face and lip detector
are off-the-shelf classifiers. Besides choosing datasets we
must also define the training parameters. Table 1 presents
the parameters used for training of the cascade classifier.
The success of the approach is related to the performance
of the classifier itself. By defining a high number of stages
we are creating several stages that the images must overcome
to be considered a positive example. The high true
positive rate ensures that almost every positive example is
Figure 6: Examples of images used to train a cascade classi-
fier for leaf detection. On the top row the original image, on
the bottom row the croped example used for training.
learned per stage. The max false positive rate creates some
margin for error, allowing the training to achieve the minimum
true positive rate per stage and a low positive rate at
the end of the cascade. Similar parameters were used and
discussed in (Lienhart, Kuranov, and Pisarevsky 2003).
Once the classifiers are obtained, they are used to assign
fitness in the course of the evolutionary runs in an attempt
to find images that are recognized as faces, lips, breasts and
leaves. We performed 30 independent evolutionary runs for
each of these classes. In summary we have 4 classifiers, with
30 independent EC runs, totaling 120 EC runs.
The settings of the GP engine, presented in table 2, are
similar to those used in previous experimentation in different
problem domains. Since the classifiers used only deal with
greyscale information, the GP engine was also limited to the
generation of greyscale images. The population size used
in this experiments 100 while in previous experiments we
used a population size of 50 (Machado, Correia, and Romero
2012a). This allows us to sample a larger portion of the
search space, contributing to the discovery of images that fit
the positive class.
In all evolutionary runs the GP engine was able to evolve
images classified as the respective objects. Similarly to
the behavior reported by Machado, Correia, and Romero
(2012a; 2012a), the GP engine was able to exploit weaknesses
of the classifier, that is, the evolved images are classi-
fied as the object but, from a human perspective, they often
fail to resemble the object. In figure 8 we present examples
of such failures. As it can be observed, it is hard to
recognize breasts, faces, leafs or lips in the presented images.
It is important to notice that these weaknesses are not
a byproduct of the fitness assignment scheme, as such they
cannot be solved by using a different fitness function, nor
particular to the classifiers used. Although different classiProceedings
of the Fourth International Conference on Computational Creativity 2013 27
Figure 7: Examples of images belonging to the negative
dataset used for training the cascade classifiers.
Table 2: Parameters of the GP engine. See (Machado and
Cardoso 2002) for a detailed description.
Parameter Setting
Population size 100
Number of generations 100
Crossover probability 0.8 (per individual)
Mutation probability 0.05 (per node)
Mutation operators sub-tree swap, sub-tree
replacement, node insertion,
node deletion, node mutation
Initialization method ramped half-and-half
Initial maximum depth 5
Mutation max tree depth 3
Function set +, !, ⇥ , /, min, max, abs,
neg, warp, sign, sqrt,
pow, mdist, sin, cos, if
Terminal set x, y, random constants
fiers have different weaknesses, we confirmed that several of
the evolved images that do not resemble faces are also recognized
as faces by commercially available and widely used
classifiers.
These results have opened a series of possibilities, including
the use of this approach to assess the robustness of object
detection systems, and also the use of evolved images as part
of the training set of these classifiers in order to overcome
some of their shortcomings. Although we already are pursuing
that line of research and promising results have been
obtained (Machado, Correia, and Romero 2012b), it is beyond
the scope of the current paper.
When one builds a face detector, for instance, one is typically
interested in building one that recognizes faces of all
types, sizes, colors, sexes, in different lighting conditions,
against clear and cluttered backgrounds, etc. Although the
inclusion of all these examples may lead to a robust clas-
(a) (b)
(c) (d)
Figure 8: Examples of evolved images identified as objects
by the classifiers that do not resemble the corresponding objects
from a human perspective. This images were recognized
as breasts (a), faces (b), leafs (c) and lips (d).
sifier that is able to detect all faces present in an image, it
will also means that this classifier will be prone to recognize
faces even when only relatively few features are present. In
contrast, when building classifiers for the purpose described
in this paper, one may select for positive examples clear and
iconic images. Such classifiers would probably fail to identify
a large portion of real-world images containing the object.
However, they are would be extremely selective and,
as such, the evolutionary runs would tend to converge to images
that clearly match the desired object. Thus, although
this was not explored, building a selective classifier can significantly
reduce the number of runs that converge to atypical
images such as the ones depicted in figure 8.
According to our subjective assessment, some runs were
able to find images that actually resemble the object that we
are trying to evolve. These add up to 6 runs from the face
detector, 5 for the lip detector, 4 for the breast detector and
4 for the leaf detector.
In figures 9,10, 11 and 12 we show, according to our
subjective assessment, some of the most interesting images
evolved. These results allow us to state that, at least in some
instances, the GP engine was able to create figurative images
evocative of the objects that the object detector was design
to recognize as belonging to the positive class.
By looking at the faces, figure 9, we can observe the presence
of at least 3 facial features per image (such as eyes,
lips, nose and head contour). The images from the first row
have been identified by users as resembling wolverine. The
Proceedings of the Fourth International Conference on Computational Creativity 2013 28
Figure 9: Examples of some of the most interesting images
that have been evolved using face detection to assign fitness.
ones of the second row, particularly the one on the left, have
been identified as masks (more specifically african masks).
In what concerns the images from the last row, we believe
that their resemblance “ghost-like” cartoons is striking.
In what concerns the images resulting from the runs where
a lip detector was used to assign fitness, we consider that
their resemblance with lips, caricatures of lips, or lip logos,
is self evident. The iconic nature of the images from the last
row is particularly appealing to us.
The results obtained with the breast detector reveal images
with well-defined or exaggerated features. We found
little variety in these runs, with changes occurring mostly
at the pixel intensity and contrast level. As previously mentioned,
most of these runs resulted in unrecognizable images
(see figure 8), which is surprising since the nature of the
function set would lead us to believe that it should be relatively
easy to evolve such images. Nevertheless, the successful
runs present images that are clearly evocative of breasts.
Finally the images from the leaf detector, vary in type and
shape. They share however a common feature they tend to
be minimalist, resembling logos. In each of the images of
the first row the detector identified two leaf shapes. On the
Figure 10: Examples of some of the most interesting images
that have been evolved using a detector of lips to assign
fitness.
Figure 11: Examples of some of the most interesting images
that have been evolved using a detector of breasts to assign
fitness.
Proceedings of the Fourth International Conference on Computational Creativity 2013 29
Figure 12: Examples of some of the most interesting images
that have been evolved using a detector of leafs to assign
fitness.
others a single leaf shape was detected.
In general, when the runs successfully evolve images that
actually resemble the desired object, they tend to generate
images that exaggerate the key features of the class. This
is entirely consistent with the fitness assignment scheme
that values images that are recognized with a high degree
of certainty. This constitutes a valuable side effect of the
approach, since the evolution of caricatures and logos fits
our intention to further explore these images from a artistic
and design perspective. The convergence to iconic, exaggerated
instances of the class, may indicate the occurrence of
the “Peak Shift Principle”, but further testing is necessary to
confirm this interpretation of the results.
Conclusions
The goal of this paper was to evolve different figurative images
by evolutionary means, using a general-purpose expression
based GP image generation engine and object detectors.
Using the framework presented by Machado, Correia,
and Romero (2012a), several object detectors were used to
evolve images that resemble: faces, lips, breasts and leafs.
The results from 30 independent runs per each classifier
shown that is possible to evolve images that are detected as
the corresponding objects and that also resemble that object
from a human perspective. The images tend to depict an
exaggeration of the key features of the associated object, allowing
the exploration of these images in design and artistic
contexts.
The paper makes 3 main contributions, addressing: (i) A
well-known open problem in evolutionary art; (ii) The evolution
of figurative images using a general-purpose expression
based EC system; (iii) The generalization of previous
results.
The open problem of finding a compact symbolic expression
that matches a target image is addressed by generalization:
instead of trying to match a target image we evolve
individuals that match a given class. Previous results (see
(Machado, Correia, and Romero 2012a)) concerned only the
evolution of faces. Here we demonstrate that other classes
of objects can be evolved. As far as we know, this is the
first autonomous system that proved able to evolve different
types of figurative images. Furthermore the experimental
results show that this is attainable with off-the-shelf and
purpose build classifiers, demonstrating that the approach is
both generalizable and customizable.
Currently, we are performing additional tests with different
object detectors in order to expand the types of imagery
produced.
The next steps will comprise the following: combine, re-
fine and explore the evolved images, using them in userguided
evolution and automatic fitness assignment schemes;
combine multiple object detectors to help refine the evolved
images (for instance use a face detector first and an eye or a
lip detector next); use the evolved examples that are seen as
shortcomings of the classifier to refine the training set and
boost the existing detectors.
Acknowledgements
This research is partially funded by: the Portuguese Foundation
for Science and Technology in the scope of project
SBIRC (PTDC/EIA–EIA/115667/2009) and of the iCIS
project (CENTRO-07-ST24-FEDER-002003), which is co-
financed by QREN, in the scope of the Mais Centro Program
and European Union’s FEDER; Xunta de Galicia Project
XUGA?PGIDIT10TIC105008PR.
<references_biblio/>
References
Baker, E. 1993. Evolving line drawings. Technical Report
TR-21-93, Harvard University Center for Research in Computing
Technology.
Baluja, S.; Pomerlau, D.; and Todd, J. 1994. Towards automated
artificial evolution for computer-generated images.
Connection Science 6(2):325–354.
DiPaola, S. R., and Gabora, L. 2009. Incorporating characteristics
of human creativity into an evolutionary art algorithm.
Genetic Programming and Evolvable Machines
10(2):97–110.
Freund, Y., and Schapire, R. E. 1995. A decision-theoretic
generalization of on-line learning and an application to
Proceedings of the Fourth International Conference on Computational Creativity 2013 30
boosting. In Proceedings of the Second European Conference
on Computational Learning Theory, EuroCOLT ’95,
23–37. London, UK, UK: Springer-Verlag.
Frowd, C. D.; Hancock, P. J. B.; and Carson, D. 2004.
EvoFIT: A holistic, evolutionary facial imaging technique
for creating composites. ACM Transactions on Applied Perception
1(1):19–39.
Griffin, G.; Holub, A.; and Perona, P. 2007. Caltech-256
object category dataset. Technical Report 7694, California
Institute of Technology.
Johnston, V. S., and Caldwell, C. 1997. Tracking a criminal
suspect through face space with a genetic algorithm. In
Back, T.; Fogel, D. B.; and Michalewicz, Z., eds., ¨ Handbook
of Evolutionary Computation. Bristol, New York: Institute
of Physics Publishing and Oxford University Press.
G8.3:1–8.
Lewis, M. 2007. Evolutionary visual art and design. In
Romero, J., and Machado, P., eds., The Art of Artificial
Evolution: A Handbook on Evolutionary Art and Music.
Springer Berlin Heidelberg. 3–37.
Lienhart, R., and Maydt, J. 2002. An extended set of haarlike
features for rapid object detection. In International Conference
on Image Processing, volume 1, I–900 – I–903 vol.1.
Lienhart, R.; Kuranov, E.; and Pisarevsky, V. 2003. Empirical
analysis of detection cascades of boosted classifiers for
rapid object detection. In DAGM 25th Pattern Recognition
Symposium, 297–304.
Machado, P., and Cardoso, A. 2002. All the truth about
NEvAr. Applied Intelligence, Special Issue on Creative Systems
16(2):101–119.
Machado, P., and Romero, J. 2011. On evolutionary
computer-generated art. The Evolutionary Review: Art, Science,
Culture 2(1):156–170.
Machado, P.; Correia, J.; and Romero, J. 2012a. Expressionbased
evolution of faces. In Evolutionary and Biologically
Inspired Music, Sound, Art and Design - First International
Conference, EvoMUSART 2012, Malaga, Spain, April 11- ´
13, 2012. Proceedings, volume 7247 of Lecture Notes in
Computer Science, 187–198. Springer.
Machado, P.; Correia, J.; and Romero, J. 2012b. Improving
face detection. In Moraglio, A.; Silva, S.; Krawiec, K.;
Machado, P.; and Cotta, C., eds., Genetic Programming -
15th European Conference, EuroGP 2012, Malaga, Spain, ´
April 11-13, 2012. Proceedings, volume 7244 of Lecture
Notes in Computer Science, 73–84. Springer.
Machado, P.; Romero, J.; and Manaris, B. 2007. Experiments
in computational aesthetics: An iterative approach
to stylistic change in evolutionary art. In Romero, J., and
Machado, P., eds., The Art of Artificial Evolution: A Handbook
on Evolutionary Art and Music. Springer Berlin Heidelberg.
381–415.
McCormack, J. 2005. Open problems in evolutionary music
and art. In Rothlauf, F.; Branke, J.; Cagnoni, S.; Corne,
D. W.; Drechsler, R.; Jin, Y.; Machado, P.; Marchiori, E.;
Romero, J.; Smith, G. D.; and Squillero, G., eds., EvoWorkshops,
volume 3449 of Lecture Notes in Computer Science,
428–436. Springer.
McCormack, J. 2007. Facing the future: Evolutionary possibilities
for human-machine creativity. In Romero, J., and
Machado, P., eds., The Art of Artificial Evolution: A Handbook
on Evolutionary Art and Music. Springer Berlin Heidelberg.
417–451.
Nishio, K.; Murakami, M.; Mizutani, E.; and N., H. 1997.
Fuzzy fitness assignment in an interactive genetic algorithm
for a cartoon face search. In Sanchez, E.; Shibata, T.; and
Zadeh, L. A., eds., Genetic Algorithms and Fuzzy Logic Systems:
Soft Computing Perspectives, volume 7. World Scientific.
Norton, D.; Darrell, H.; and Ventura, D. 2010. Establishing
appreciation in a creative system. In Proceedings of the First
International Conference Computational Creativity, 26–35.
Papageorgiou, C. P.; Oren, M.; and Poggio, T. 1998. A general
framework for object detection. In Sixth International
Conference on Computer Vision, 555–562.
Romero, J.; Machado, P.; Santos, A.; and Cardoso, A. 2003.
On the development of critics in evolutionary computation
artists. In Gunther, R., et al., eds., ¨ Applications of Evolutionary
Computing, EvoWorkshops 2003: EvoBIO, EvoCOMNET,
EvoHOT, EvoIASP, EvoMUSART, EvoSTOC, volume
2611 of LNCS. Essex, UK: Springer.
Santana, M. C.; Deniz-Su ´ arez, O.; Ant ´ on-Canal ´ ´ıs, L.; and
Lorenzo-Navarro, J. 2008. Face and facial feature detection
evaluation - performance evaluation of public domain
haar detectors for face and facial feature detection. In Ranchordas,
A., and Araujo, H., eds., ´ VISAPP (2), 167–172.
INSTICC - Institute for Systems and Technologies of Information,
Control and Communication.
Saunders, R., and Gero, J. 2001. The digital clockwork
muse: A computational model of aesthetic evolution. In
Wiggins, G., ed., AISB’01 Symposium on Artificial Intelligence
and Creativity in Arts and Science, 12–21.
Secretan, J.; Beato, N.; D’Ambrosio, D. B.; Rodriguez, A.;
Campbell, A.; Folsom-Kovarik, J. T.; and Stanley, K. O.
2011. Picbreeder: A case study in collaborative evolutionary
exploration of design space. Evolutionary Computation
19(3):373–403.
Sims, K. 1991. Artificial evolution for computer graphics.
ACM Computer Graphics 25:319–328.
Ventrella, J. 2010. Self portraits with mandelbrot genetics.
In Proceedings of the 10th international conference
on Smart graphics, SG’10, 273–276. Berlin, Heidelberg:
Springer-Verlag.
Viola, P., and Jones, M. 2001. Rapid object detection using
a boosted cascade of simple features. Computer Vision and
Pattern Recognition, IEEE Computer Society Conference on
1:511.
World, L. 1996. Aesthetic selection: The evolutionary art of
steven Rooke. IEEE Computer Graphics and Applications
16(1).
Proceedings of the Fourth International Conference on Computational Creativity 2013 31