Conference PaperPDF Available

Evotype: Evolutionary Type Design

Authors:

Abstract and Figures

An evolutionary generative system for type design, Evotype, is described. The system uses a Genetic Algorithm to evolve a set of individuals composed of line segments, each encoding the shape of a specific character, i.e. a glyph. To simultaneously evolve glyphs for the entire alphabet, an island model is adopted. To assign fitness we resort to a scheme based on Optical Character Recognition. We study the evolvability of the proposed approach as well as the impact of the migration in the evolutionary process. The migration mechanism is explored through three experimental setups: fitness guided migration, random migration, and no migration. We analyse the experimental results in terms of fitness, migration paths, and appearance of the glyphs. The results show the ability of the system to find suitable glyphs and the impact of the migration strategy in the evolutionary process.
Content may be subject to copyright.
Evotype: Evolutionary Type Design
Tiago Martins, Jo˜ao Correia, Ernesto Costa, and Penousal Machado
CISUC, Department of Informatics Engineering,
University of Coimbra, 3030 Coimbra, Portugal
{tiagofm,jncor,ernesto,machado}@dei.uc.pt
Abstract. An evolutionary generative system for type design, Evotype,
is described. The system uses a Genetic Algorithm to evolve a set of
individuals composed of line segments, each encoding the shape of a
specific character, i.e. a glyph. To simultaneously evolve glyphs for the
entire alphabet, an island model is adopted. To assign fitness we resort to
a scheme based on Optical Character Recognition. We study the evolv-
ability of the proposed approach as well as the impact of the migration in
the evolutionary process. The migration mechanism is explored through
three experimental setups: fitness guided migration, random migration,
and no migration. We analyse the experimental results in terms of fit-
ness, migration paths, and appearance of the glyphs. The results show
the ability of the system to find suitable glyphs and the impact of the
migration strategy in the evolutionary process.
Keywords: type design, evolutionary design, island model.
1 Introduction
Although conventional computational design tools are eective for precise design
tasks during the later phases of the design process, they oer insucient support
to design exploration during the earliest, essentially conceptual, stages of the
design process.
We present an evolutionary system for type design — Evotype. Although it
is still a work in progress, the system is already able to automatically generate
alternative designs for glyphs from scratch. A glyph consists in a specific graphic
expression of a given readable character. For the purposes of this article, we focus
in the evolution of glyphs for letters of the Roman alphabet.
The main contribution presented herein is a functional prototype of a gener-
ative system capable of creating consistent glyphs. Other contributions include:
(i) a Genetic Algorithm (GA) with a generic representation wherein individuals
are composed by line segments encoded as sequence of numeric values; (ii) a
fully autonomous evolutionary approach for the evolution of glyphs; (iii) the use
of an island model and the study of the impact of migration on evolution; and
(iv) the use of a Machine Learning (ML) approach to guide the evolution and
the migration process.
The paper is organised as follows: section 2 presents related work, considering
applications of evolutionary techniques in the domain of glyphs design; section 3
thoroughly describes the GA of the proposed system; section 4 describes the
experimental setup; section 5 presents the analysis of the experimental results;
finally, conclusions and future work is presented in section 6.
2 Related Work
Designers, engineers, artists, and scientists have been using evolution-based tech-
niques to support the creative process and evolve innovative artefacts. Although
evolutionary glyph design is a relatively unexplored area, some applications exist.
Butterfield and Lewis [1] employ Interactive Evolutionary Computation (IEC)
to explore the creation of fonts. More specifically, they evolve deformations, i.e.
the letters of a specific typeface are deformed by a set of implicit surface primi-
tives, which are encoded in the genotypes. Lund [2] also uses IEC to evolve the
settings of a parametric typeface system. Each parameter controls the appear-
ance of a given characteristic of the font. Unemi and Soda [3] propose an IEC
system for the design of Japanese Katakana glyphs. Schmitz [4] presents the
interactive program genoTyp, which allows the user to create new fonts through
the breeding of existing ones, according to genetic rules and manual manipula-
tion of their genes. The possibility of recombining famous typefaces is exciting,
however, the limitations of the representation hinder the quality of the results.
Levin et al. [5] use IEC to implement the Alphabet Synthesis Machine, a system
which allows the creation and evolution of abstract letter forms.
Despite evolutionary systems for glyph design exist, as far as we know all
of them rely on user evaluation. Thus, the user interactively iterates a cyclic
process of selection and generation until an acceptable solution is obtained. As
such, all suer from the well-known limitations of IEC systems, namely user-
fatigue and inconsistency in evaluation. Additionally, they require the creation
of a parametric typeface (e.g., [2]), or pre-existing typefaces or skeletons (e.g.,
[4] and [1], respectively), and are conditioned by these requirements.
3 The Approach
Evotype evolves glyph designs for various characters in a parallel and autonomous
way. To achieve this, a GA [6] is implemented to evolve dierent populations of
candidate glyph designs. Each individual is a glyph design. Each population lives
in its own island and is composed of individuals that represent a specific charac-
ter. Thus, to evolve glyphs for 26 characters we use 26 populations in 26 islands.
The dierent islands can communicate with each other, allowing the migration
of glyphs among them.
The system is schematically represented in figure 1 and behaves as follows.
The evolutionary process begins with the initialisation of all populations with
randomly created glyphs. The individuals are evaluated and then selected for
mating according to their fitness. Recombination and mutation operators are
applied to generate ospring. The selection stage follows, determining which
individuals proceed to the next generation. The next step is migration, where
C
B
AISLAND
INITIALISATION EVALUATION SELECTION
RECOMBINATION
MUTATION
EVALUATION
SELECTION
MIGRATION
TERMINATION
!
!
Fig. 1. Overview of Evotype.
individuals may be moved to or received from other islands. As depicted in
figure 1 the process is cyclically repeated. The termination criterion is based on
the number of generations. In the following subsections we detail the genetic
mechanisms employed in Evotype.
3.1 Representation
The genotype consists of a sequence of genes encoding a glyph. Each gene codifies
a two-dimensional line segment that is composed by a sequence of five numbers,
wherein the first four are the coordinates of its end points and last one correspond
to its thickness (see figure 2). The genotype’ length may vary from individual
to individual, thus dierent individuals can be composed of dierent number of
line segments.
A mapping mechanism, normally referred to as embryogenesis, is in charge for
the expression of the genotype into a perceptible artefact—the phenotype. The
phenotype consists in a graphical representation of the genotype, i.e. a glyph
created from the encoded parameters. The expression process consists in the
drawing of dark line segments, defined by the genotype, on a white canvas, as
illustrated in figure 3.
{ [X1, Y1,X2,Y2 ,T ], [X1, Y1,X2,Y2, T ],...}
1
2
3
4
1
2
3
4
Fig. 2. Encoding of the genotype, composed of genes that codify line segments (1)
defined by their end points (2and 3)theirthickness(4).
The dimension of the search space is reduced through the use of a rect-
angular grid that constraints the coordinates of the line segments’ end points,
which must adhere to the grid points. The density of the grid is configured by
{[0,0,2,2,2],
[4,0,2,2,2],
[0,0,0,4,4],
[2,2,2,3,4]
[4,0,4,4,4]}
12340
1
2
3
4
0
1
2
3
4
5
1
2
4
5
3
Fig. 3. Mapping process from genotype to phenotype. On the left, the genotype; On
the middle, an intermediate representation depicting the grid and the correspondence
between genes and line segments; On the right, the phenotype.
the user.The representation ensures high locality. Neighbouring genotypes are
mapped to similar phenotypes, meaning that small modifications of the genetic
code induce small changes in the phenotypic space.
3.2 Initialisation
The initial populations are seeded with randomly generated glyphs. Each glyph
of the first population is composed of a single line segment, with all the gene
values set by uniform random selection over the admissible interval for each
parameter. All islands receive identical initial populations. This initialisation
setup provides equality and simplicity among all initial populations, enabling
us to access the ability of the system to evolve glyphs for dierent characters,
starting from the same set of random glyphs.
3.3 Crossover
The crossover operation consists in the exchange of line segments between par-
ents. Crossover operates on gene boundaries, preserving the integrity of line
segments. The operator proceeds as follows: randomly select a rectangular area
of the grid; determine, for both parents, the line segments whose middle points
are inside the rectangle; exchange those line segments between parents. As illus-
trated in figure 4, crossover may be asymmetric, in the sense that the number of
line segments a genotype “receives” may be dierent from the one it “donates”.
3.4 Mutation
Mutation also operates on a gene basis. Thus, the mutation of a gene implies
changing one of its five parameters by a value of one, as illustrated in figure 5.
At the phenotype level, this variation results in the minimum translation of one
of the end points of the line segment in one of the four possible directions—up,
down, left, or right—or the minimum variation of its thickness. The impact of
this change at the phenotype level depends on the density of the grid. A denser
grid allows smaller visual variations. The probability of mutation is defined per
gene, meaning that multiple genes may be mutated.
A second type of mutation exists, gene deletion and insertion, allowing the
variation of the size of the genotype. There is a probability of deleting a randomly
selecting gene and inserting a randomly created one.
{[2,0,4,4,4],
[0,2,1,4,4]}
{[0,2,1,0,4],
[1,0,3,0,4],
[0,2,1,4,4]}
[4,2,3,4,2],
[2,2,4,2,2],
{[0,2,1,0,4],
[4,2,3,4,2]}
[0,4,3,2,2],
[0,4,2,0,4],
[2,2,4,2,2],
[1,0,3,0,4],
[0,4,3,2,2]}
{[2,0,4,4,4],
[0,4,2,0,4],
1
1
3
3
2
2
1
2
3
2
3
1
Fig. 4. Crossover process. On the left, two parents, the corresponding genotypes, and
a randomly selected rectangular area (shaded region). The selected area determines
the line segments that will be exchanged (1, 2, and 3). On the right, the results of the
crossover operation at the genotype and phenotype level.
[0,4,2,0,4]}
{[0,4,3,3,2],
[2,0,4,4,4],
[0,4,2,0,4]}
{[0,4,3,2,2],
[2,0,4,4,4],
Fig. 5. Mutation process. On the left, the original genotype and phenotype; On the
right, the results of the mutation operator.
Unfeasible variations are prevented during the mutation, including transla-
tions of coordinates that (i) go beyond the grid limits, (ii) create line segments
with null length, or (iii) generate two line segments which are defined by the
same end points.
3.5 Evaluation
As previously mentioned, each population (i.e. each island) is composed of in-
dividuals that are candidate graphic representations of a specific character. As
such, the fitness of an individual depends on the environment, i.e. the island were
it lives. In the scope of this paper we use Optical Character Recognition (OCR)
to assign fitness. The details of the fitness assignment scheme are described in
section 4.1.
3.6 Migration
Migration can occur once per each island in each generation. The probability
of occurrence is determined by the migration rate. To study the influence of
migration in the evolutionary process we considered three migration mechanisms:
No Migration – As the name indicates, no migration is used. The islands are
isolated.
Random – When migration occurs, each island selects one random individual
among the ones living on dierent islands. A copy of the individual is added
to the population of the island, replacing the individual with worst fitness.
Fig. 6. Screenshot of the graphic user interface of Evotype. A demo video can be seen
on http://cdv.dei.uc.pt/2015/evotype.mov.
Fitness Guided – When migration occurs, each island evaluates all its indi-
viduals according to the environment of the other islands. A copy of the
individual that attains the highest fitness is added to the queue of that des-
tination island. When this process is completed for all islands, each island
checks its immigration queue, which may be empty, and selects the fittest
individual. This individual is added to the population of the island, replacing
the individual with worst fitness.
It is important to remember that in Evotype fitness is local. Therefore, the
fitness of an individual depends on the environment (i.e. the character it is trying
to represent graphically) and each island corresponds to a dierent character.
3.7 Visualisation
Evotype is able to fluidly show the evolutionary process through a simple graphic
user interface, conceived to allow the user to visualise the dierent evolving
glyphs of all islands (see figure 6). Islands (each corresponding to a character)
are arranged horizontally in dierent columns, in alphabetical order from left
to right. The individuals (glyphs) of the current generation of each island are
depicted vertically, in descending order of fitness. At any moment of the evolu-
tionary process, the user can export the evolved glyphs as vectors files to make
further design refinements.
The glyphs’ fitness is visualised through a simple graphic approach. The
fitness value of each individual is represented through a horizontal line that
overlays the corresponding column, and is vertically positioned according to the
mapped value of the fitness value to the height of the interface (higher fitness
values on top). We consider this fitness visualisation technique functional for the
purposes of this work, in the sense that we are not particularly interested in
seeing the specific fitness values. Instead, we wish to visualise the distribution
and convergence of fitness over time, and the comparison of fitness values among
islands.
Tab l e 1 . Evolutionary System
Parameter Setting
Number of runs 30
Number of generations 100
Population size 100
Crossover rate 0.8
Delete gene mutation rate 0.075
Insert gene mutation rate 0.075
Change gene mutation rate 0.3
Selection method Tournament
Tournament size 3
Elite size 1
Glyph grid size 40 40
Tab l e 2 . Classifier
Parameter Setting
Input image size 48 48
Quantized colours 5
Threshold ()200
Promotion () 1.005
Demotion () 0.995
Initial weights values 2
Training iterations 1000
Examples per island 78
4 Experimental Setup
In this work, we evolve glyphs for all the uppercase letters of the Roman alpha-
bet, so a total of 26 islands are considered. We conduct experiments to assess
the adequacy of the engine for the evolution of glyphs. Furthermore, we study
the impact of the migration policy (see section 3.6) in evolution.
The experimental parameters used in the course of the experiments described
in this paper are summarised in tables 1 and 2. In the following subsections we
detail the fitness assignment scheme for the experiments.
4.1 Fitness
In this work, we have 26 islands, each one evolving towards a dierent objective,
so we need to provide a proper fitness function to guide evolution for each island.
We consider that interactive evolution of all these islands’ populations would be
an hard and tedious task. For that reason, we sought to use an automatic fitness
assignment scheme.
We are dealing with the evolution of character glyphs. Certainly, one of the
preconditions of a visual character representation is its recognisability. As such,
we evolve images that are recognised automatically as specific characters. We
choose an OCR ML approach to automatically assign fitness. Although the use
of ML to assign fitness is not novel (refer to, e.g., [7,8,9]).
Each individual is processed as a bit map image by an OCR system, and
the intermediary values from OCR process are used to assign fitness. Based on
the work of Burry et al. [10] we use Sparse Network of Winnow (SNoW) [11], a
sparse network of linear units, as the classifier system for the OCR approach. The
SNoW classifier is characterised as having two layers, the input layer and ntarget
nodes, which are linked through weighted edges. To perform OCR, we train,
oine and per character, 26 dierent classifiers. The input examples consist of
bitmap images corresponding to the characters of 78 dierent typefaces. These
are pre-processed to extract features, which are used in training/classification.
STANDARISED FITNESS
ABCDEFGHIJKLMNOPQRSTUVWXYZ
1.00
0.95
0.90
0.85
Fig. 7. Standardised fitness of the best individual of the last generation of each island.
With a solid line, the no migration; dashed line, the random migration; and dotted
line, the fitness guided migration. The visualised results are averages of 30 runs.
The process and decisions made for the feature extraction are based on the work
of Burry et al. [10]. Table 2 summarises the overall OCR system parameters.
To deal with the OCR multi-class classification problem we use a one versus
all strategy for training and classification. We train each classifier by treating all
the instances of the character as the positive class and all instances of all other
characters as the negative class. Thus, each classifier has two target nodes, one
for the positive class (i.e. the character in question) one for the negative class
(i.e. all the remaining characters). The activations of the nodes indicates the
degree of membership of the input image to the respective class.
To assign fitness each input image is mapped to a 2Dspace where the x
coordinate corresponds to the activation value of the positive class node of the
classifier, and the ycoordinate corresponds to the activation value of the negative
class node. Thus, the ideal scenario would be maximising xwhile minimising y.
Unfortunately, this is typically not possible since the letters share characteristics
among them. For instance, the input pattern that maximizes the xcoordinate
for Q, will, necessarily, yield a high yvalue due to the presence of examples of
the letter O in the negative class. Therefore, as is often the case multi-objective
optimization problems, a compromise between xand yis necessary to obtain
good results.
As such, we adopted the following procedure, establish a target activation
point empirically. We begin by calculating the input pattern that minimises the y
coordinate, the activation value of the output node corresponding to the negative
class for this pattern becomes yt. Then, also analytically, we determine the input
pattern for which the absolute dierence between the activation values between x
and yis minimal. Such input, typically, has the features necessary for the image
to be classified as a member of the positive class, but still possesses features
that are common with other characters. The activation value of the output node
corresponding to the positive class for this pattern becomes xt. Finally the fitness
of a new input image is based on the following formula: dist(x)=k(xt,y
t)
(xi,y
i)k,wherexi,yiare the activation values of the output nodes associated
ABCDEFGHIJKLMNOPQRSTUVWXYZ
100
80
60
40
20
0
SETUPS GROUPED PER ISLAND
GENERATION
Fig. 8. Progression of the fitness of the fittest individual of each island over 100 gen-
erations. The results are grouped per island from left to right in the following order:
no migration,random migration and fitness guided migration. Each graphic bar repre-
sents the variation of the standardise fitness of the best individual from 0 to 1 across
the generations, divided in 5 dierent intervals of equal size (0.2). These intervals are
represented by dierent shades of grey which change from darker to lighter according
to its fitness value (from lower to higher). The visualised results are averages of 30
runs.
with the input image, and results in the euclidean distance between these two
points. Since we wish to minimise the distance to this point, fitness becomes:
fitness(x)=1/(1 + dist(x)).
5 Experimental Results
We begin the analysis of the experimental results by focusing on the fitness
values obtained by the 3 migration strategies. Figure 7 summarizes these results
by presenting the average fitness of the best individual of the last generation.
For the purposes of readability the fitness values have been normalized to [0,1],
by dividing the raw fitness scores by the maximum fitness score obtained for
each character in the course of the 90 runs (30 per migration strategy). A brief
perusal of the results indicates that fitness based migration outperforms the
other two strategies, attaining higher fitness values for 20 out of the 26 islands.
There are two “ties” among strategies, namely for islands I and N. Random
migration outperforms the other methods for three islands (A, W, and Z), while
no migration obtains the best results for island T.
To better understand the dynamics of the evolutionary process we summarise
in figure 8 the evolution of fitness over time. The results suggest that the use
of migration (fitness guided or ra ndo m ) promotes a faster convergence for max-
imum fitness values. To illustrate how the evolution of fitness is reflected in the
graphic appearance of the glyphs, figure 9 depicts the glyphs evolved through
time in a typical run using fitness based migration.
0
4
8
12
16
20
50
75
99
GENERATION
Fig. 9. Best individuals of each island for dierent generations from a typical run using
fitness guided migration.
Regarding the speed of convergence to maximum fitness values, fitness guided
is the first approach to attain high fitness values in 22 of 26 islands, with the
exceptions being the letters A, V, W, and Z. These exceptions can be justified
by the analysis of the migration paths along the evolutionary process. Figure 10
depicts the average number of migrations among islands for fitness guided mi-
gration.
The results in these islands can be justified as follows. For the letter Z,
the migration paths indicate that the island does not receive immigrants from
other islands, as such the performance of the no migration and fitness guided
migration should be, and is, comparable. The same justification explains the
results obtained for island A and V. The explanation for the behaviour observed
in island W is harder to explain, in this fitness guided case migration appears
to be detrimental to the evolutionary process, the migration of individuals from
island A to island W, although helpful in the beginning of the run, leads to
premature convergence to sub-optimal solutions.
Figure 10 also shows that islands corresponding to visually similar characters
migrate more individuals among them. Some examples of these migration paths
are the following: A!W; B!D; E!B,F; F!P; K,X!N; L!E; H,N,V!M;
CG; DO; IT; and PR.
Although the analysis of the evolutionary process is valuable, from an evolu-
tionary design point of view, the analysis of the visual output of the system is
also important. The visual results presented in figure 11 highlight the diversity of
the glyphs evolved in dierent runs. Additionally, and although this is dicult to
measure objectively, migration appears promotes the visual coherence among the
characters of dierent islands, which is important from a typeface design point
of view. These outputs may be provided to a type designer as alternative sources
of inspiration, thus arguably assisting creation of glyphs during the conceptual
phase of the creative process. The analysis of the impact of these suggestions is
outside the scope of the paper, and will be left for a future opportunity.
2
1
18
1
8
1
2
1
2
3
4
4
5
4
15
1
1
1
17
1
1
8
2
5
11
4
2
1
18
9
1
1
1
17
1
7
4
7
4
1
3
1
9
3
1
2
3
1
3
45
3
2
1
1
8
3
1
3
1
8
3
1
2
7
6
1
1
2
5
1
1
13
7
1
2
11
4
5
1
1
1
2
1
6
4
22
1
19
3
25
1
6
1
7
17
8
1
3
6
35
2
21
8
7
3
3
24
1
9
4
1
3
15
5
2
2
1
4
3
1
29
1
5
1
3
37
6
9
1
2
1
5
16
1 1
1
7
2
1
1
A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
U
V
W
X
Y
Z
ABCDEFGHIJKLMNOPQRSTUVWXYZ
DESTINATION
ORIGIN
34
37
39
38
44
42
38
30
43
37
22
32
19
23
40
39
26
31
32
45
24
31
23
43
32
29
2 29 43 46 32 31 36 28 51 69 39 36 77 78 41 55 41 39 9 37 18 6 17 11 2
Fig. 10. Visualisation of the average number of migrations between the dierent islands
when using fitness guided migration. Origin islands are positioned horizontally and the
destiny islands vertically. High numbers of migrations are highlighted by darker squares.
The results are averages of 30 runs.
6 Conclusions and Future Work
We have presented Evotype, an evolutionary approach for the automatic gen-
eration of character glyphs with an automatic fitness assignment scheme based
on an OCR approach. The approach adopts an island model where the glyphs
corresponding to each character populate each island. The experimental results
shows that migration of individuals among islands is beneficial, provided that
the fitness in the destination environment is taken into consideration. More im-
portantly, the experimental results show that Evotype provides an ecient ar-
chitecture to evolve and explore alternatives for glyph design.
Future work will focus on: (i) the extension of the genetic representation to
allow a wider range of graphic primitives; (ii) the exploration of dierent mi-
gration policies and topologies; (iii) the exploration of other fitness assignment
schemes, which may promote the diversity, aesthetic appeal, and creative po-
tential of the glyphs; and (iv) the further development of the system as tool for
supporting the creativity of the designer.
Acknowledgments This research is partially funded by: 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; Funda¸ao para a Ciˆencia
e Tecnologia (FCT), Portugal, under the grants SFRH/BD/90968/2012 and
Fig. 11. Some examples of glyphs evolved in dierent runs.
SFRH/BD/105506/2014; project ConCreTe. The project ConCreTe acknowl-
edges the financial support of the Future and Emerging Technologies (FET)
programme within the Seventh Framework Programme for Research of the Eu-
ropean Commission, under FET grant number 611733.
References
1. Butterfield, I., Lewis, M.: Evolving fonts (2000) Consulted in http://accad.osu.
edu/~mlewis/AED/Fonts/ on October 2014.
2. Lund, A.: Evolving the shape of things to come: A comparison of direct ma-
nipulation and interactive evolutionary design. In: International Conference on
Generative Art. Domus Argenia, Rome, Italy (2000)
3. Unemi, T., Soda, M.: An iec-based support system for font design. In: Proceedings
of the IEEE International Conference on Systems, Man & Cybernetics: Washing-
ton, D.C., USA, 5–8 October 2003. (2003) 968–973
4. Schmitz, M.: genotyp, an experiment about genetic typography. Presented at
Generative Art Conference 2004 (2004)
5. Levin, G., Feinberg, J., Curtis, C.: The alphabet synthesis machine (2006) Con-
sulted in http://www.alphabetsynthesis.com on October 2014.
6. Holland, J.H.: Adaptation in Natural and Artificial Systems. University of Michi-
gan Press, Ann Arbor, Michigan (1975)
7. Baluja, S., Pomerlau, D., Todd, J.: Towards automated artificial evolution for
computer-generated images. Connection Science 6(2) (1994) 325–354
8. Machado, P., Romero, J., Manaris, B.: Experiments in computational aesthet-
ics: An iterative approach to stylistic change in evolutionary art. In Romero, J.,
Machado, P., eds.: The Art of Artificial Evolution: A Handbook on Evolutionary
Art and Music. Springer Berlin Heidelberg (2007) 381–415
9. Machado, P., Correia, J., Romero, J.: Expression-based evolution of faces. In:
Evolutionary and Biologically Inspired Music, Sound, Art and Design - First Inter-
national Conference, EvoMUSART 2012, M´alaga, Spain, April 11-13, 2012. Pro-
ceedings. Volume 7247 of Lecture Notes in Computer Science., Springer (2012)
187–198
10. Artan, Y., Burry, A., Kozitsky, V., Paul, P.: Ecient smqt features for snow-
based classification on face detection and character recognition tasks. In: Image
Processing Workshop (WNYIPW), 2012 Western New York. (Nov 2012) 45–48
11. Carlson, A., Cumby, C., Rosen, J., Roth, D.: The snow learning architecture.
Technical Report UIUCDCS-R-99-2101, UIUC Computer Science Department (5
1999)
... Therefore, we argue that it is as important to master and exploit the tools at hand, as it is to challenge those tools, by modifying them or inventing new ones that suit unique ideas and design projects. In this paper, we give a step forward of our work Evotype [11] , an evolutionary system for type design. It employs Evolutionary Computation (EC) and Machine Learning (ML) techniques to automatically generate alternative designs for glyphs (see figure 1). ...
... These parts are manually defined in an earlier stage. In Martins et al. [11] we present the first iteration of Evotype that employs a GA to evolve glyphs for the Roman alphabet using line segments. In this early iteration, we already employ an automatic fitness assignment scheme to guide the evolutionary process. ...
... Finally, parametrised approaches can overly influence their outcome, since all users start from the same base, and thus are not flexible enough. The same is observed in approaches that provide a limited range of visual elements to create the glyphs (e.g., [11]). From our analysis of the shortcomings of the related work in evolutionary type design we have extracted two main observations that guided the development of the research presented in this paper: (i) fitness assignment should be automatic , freeing the user from the need to evaluate hundreds of populations; and (ii) more than providing a final and fully functional typeface, the system should aid the designers by providing them typeface sketches that can be further refined. ...
Conference Paper
Typography plays a key communication role in the contemporary information-dense culture. Type design is a central, complex, and time consuming task. In this work we develop the generative system to type design based on an evolutionary algorithm. The key novel contributions are twofold. First, in terms of representation it relies on the use of assemblages of shapes to form glyphs. There are no limitations to the types of shapes that can be used. Second, we explore a compromise between legibility and expressiveness, testing different automatic fitness assignment schemes. The attained results show that we are able to evolve a wide variety of alternative glyphs, making the proposed system a viable alternative for real-world applications in the field of type design.
... In this paper, we continue our research on automatic evolution of glyphs [2,3]. We explore a different approach based on type stencils (see, e.g., figure 1). ...
... It employs an Interactive Genetic Algorithm (IGA) to modify the parts, presenting different options for the user to choose from in each iteration. In [2], we present Evotype, which employs a GA to evolve glyphs for the Roman alphabet using line segments. This approach uses an automatic fitness assignment scheme to guide the evolutionary process. ...
... Finally, parametrised approaches can overly influence their outcome. We observe the same issues in approaches that provide a limited range of visual elements to create the glyphs (e.g., [2]). Others lack coherence between the glyphs of the same typeface (e.g., [3]). ...
Chapter
Full-text available
Typefaces are an essential resource employed by graphic designers. The increasing demand for innovative type design work increases the need for good technological means to assist the designer in the creation of a typeface. We present an evolutionary computation approach for the generation of type stencils to draw coherent glyphs for different characters. The proposed system employs a Genetic Algorithm to evolve populations of type stencils. The evaluation of each candidate stencil uses a hill climbing algorithm to search the best configurations to draw the target glyphs. We study the interplay between legibility, coherence and expressiveness, and show how our framework can be used in practice.
... Martins et al. are developing Evotype (Martins et al., 2015(Martins et al., , 2016(Martins et al., , 2018, a system for TD that employs evolutionary computation and machine learning techniques to automatically generate designs for glyphs. In one of the project's iterations (Martins et al., 2016) the system explores the idea from a assembly a set of basic visual shapes as modules to create glyphs. ...
Conference Paper
Full-text available
Designing a modular typeface typically involves the creation of geometric relationships between shapes that are repeated in the same letterform and/or among different letterforms. One can see these relationships as workflows that follow a logic of input, processing, and output of shapes. Therefore, designing a typeface involves a step-by-step set of operations, or instructions, that enables the designer to create a typeface in an algorithmic way. This paper presents a system that allows one to design modular typefaces. The system allows the user to design a typeface by formalising a “recipe” that transforms a set of input shapes into glyphs throughout a node-based approach. The user can input a set of shapes that through a set of geometric operations are transformed and recombined. There are three types of operation nodes: move, rotate, and scale. The relationships between nodes are established by links that connect them. The output of one node is passed as an input to another creating a flow of shapes from node to node. This way, any modification in a node is immediately propagated to the following nodes and consequently to the final glyphs. To analyse the possibilities and limitations of the system, we tested the proposed system with fifteen graphic designers. The outcomes and feedback made by the users unveil the value and relevance of the system and point out future directions for this work.
... [11]) and others focus on being autonomous (e.g. Evotype [12] which generates glyph designs from scratch). ...
Conference Paper
Full-text available
The number of people who design typefaces has drastically increased in the last twenty years. However, not all typefaces work as they should, i.e., as a group of characters with shared attributes. We present a tool for helping type designers in their creative process, which explores the anatomic relations among characters of a typeface. Having computer-aided design as a goal, our tool helps in the early stages of designing a typeface by using semi-automatic letter-part sharing and allowing the users to compare their design with existing typefaces.
Article
Full-text available
Typefaces have become an essential resource used by graphic designs to communicate. Some designers opt to create their own typefaces or custom lettering that better suits each design project. This increases the demand for novelty in type design, and consequently the need for good technological means to explore new thinking and approaches in the design of typefaces. In this work, we continue our research on the automatic evolution of glyphs (letterforms or designs of characters). We present an evolutionary framework for the automatic generation of type stencils based on fitness functions designed by the user. The proposed framework comprises two modules: the evolutionary system, and the fitness function design interface. The first module, the evolutionary system, operates a Genetic Algorithm, with a novelty search mechanism, and the fitness assignment scheme. The second module, the fitness function design interface, enables the users to create fitness functions through a responsive graphical interface, by indicating the desired values and weights of a set of behavioural features, based on machine learning approaches, and morphological features. The experimental results reveal the wide variety of type stencils and glyphs that can be evolved with the presented framework and show how the design of fitness functions influences the outcomes, which are able to convey the preferences expressed by the user. The creative possibilities created with the outcomes of the presented framework are explored by using one evolved stencil in a design project. This research demonstrates how Evolutionary Computation and Machine Learning may address challenges in type design and expand the tools for the creation of typefaces.
Thesis
Full-text available
Typical Machine Learning (ML) approaches rely on a dataset and a model to solve problems. For most problems, optimisation of ML approaches is crucial to attain competitive performances. Most of the effort goes towards optimising the model by exploring new algorithms and tuning the parameters. Nevertheless, the dataset is also a key part for ML performance. Gathering, constructing and optimising a representative dataset is a hard task and a time-consuming endeavour, with no well-established guidelines to follow. In this thesis, we attest the use of Evolutionary Computation (EC) to assess and improve classifiers via synthesis of new instances. An analysis of the state of the art on dataset construction is performed. The quality of the dataset is tied to the availability of data, which in most cases is hard to control. A throughout analysis is made about Instance Selection and Instance Generation, which sheds light on relevant points for the development of our framework. The Evolutionary FramEwork for Classifier assessmenT and ImproVemEnt (EFECTIVE) is introduced and explored. The key parts of the framework are identified: the Classifier System (CS) module, which holds the ML model that is going to be assessed and improved; the EC module responsible for generating the new instances using the CS module for fitness assignment; and the Supervisor, a module responsible for managing the instances that are generated. The approach comes together in an iterative process of automatic assessment and improvement of classifiers. In a first phase, EFECTIVE is tested as a generator, creating instances of a particular class. Without loss of generality, we apply the framework in the domain of image generation. The problem that motivated the approach is presented first: frontal face generation. In this case, the framework relies on the combination of an EC engine and a CS module, i. e., a frontal face detector, to generate images of frontal faces. The results were revealing in two different ways. On the one hand, the approach was able to generate images that from a subjective standpoint resemble faces and are classified as such by the classifier. On the other hand, most of the images did not resemble faces, although they were classified as such by the classifier module. Based on the results, we extended the approach to generate other types of object, attaining similar results. We also combined several classifiers to study the evolution of ambiguous images, i. e. images that induce multistable perception. Overall, the results suggest that the framework is viable as a generator of instances and also that these instances are often misclassified by the CS module. Building on these results, in a second phase, a study of EFECTIVE for improving the performance of classifiers is performed. The core idea is to use the evolved instances that are generated by the EC engine to augment the training dataset. In this phase, the framework uses the Supervisor module to select and filter the instances that will be added to the dataset. The retraining of the classifier with these instances completes an iteration of the framework. We tested this pipeline in a face detection problem evolving instances to: (i) expand the negative dataset; (ii) expand the positive dataset; and (iii) expand both datasets in the same iteration. Overall, the results show that: expanding the negative dataset, by adding misclassified instances, reduces the number of false alarms; expanding the positive dataset increases the number of hits; expanding positive and negative datasets allows the simultaneous reduction of false alarms and increase of hits. After demonstrating the adequacy of EFECTIVE in face detection, we tested the framework in a Computational Creativity context to create an image generation system that promotes style change, obtaining results that further demonstrate the potential of the framework.
Chapter
In digital games, the visual representation of game assets such as avatars or game levels can hint at their purpose, in-game use and strengths. In the Pokémon games, this is particularly prevalent with the namesake creatures’ type and the colors in their sprites. To win these games, players choose Pokémon of the right type to counter their opponents’ strengths; this makes the visual identification of type important. In this paper, computational intelligence methods are used to learn a mapping between a Pokémon’s type and its in-game sprite, colors and shape. This mapping can be useful for a designer attempting to create new Pokémon of certain types. In this paper, instead, evolutionary algorithms are used to create new Pokémon sprites by using existing color information but recombining it into a new palette. Results show that evolution can be applied to Pokémon sprites on a local or global scale, to exert different degrees of designer control and to achieve different goals.
Chapter
The genotypic functions from apriori aesthetically evolved images are mutated progressively and their phenotypes sequenced temporally to produce animated versions. The animated versions are mapped onto typeface and combined spatially to produce animated typescript. The output is then discussed with reference to computer aided design and machine learning.
Conference Paper
Evolving images using genetic programming is a complex task and the representation of the solutions has an important impact on the performance of the system. In this paper, we present two novel representations for evolving images with genetic programming. Both these representations are based on the idea of recursively partitioning the space of an image. This idea distinguishes these representations from the ones that are currently most used in the literature. The first representation that we introduce partitions the space using rectangles, while the second one partitions using triangles. These two representations are compared to one of the most well known and frequently used expression-based representations, on five different test cases. The presented results clearly indicate the appropriateness of the proposed representations for evolving images. Also, we give experimental evidence of the fact that the proposed representations have a higher locality compared to the compared expression-based representation.
Conference Paper
Full-text available
The combination of a classifier system with an evolutionary image generation engine is explored. The framework is instantiated using an off-the-shelf face detection system and a general purpose, expression-based, genetic programming engine. By default, the classifier returns a binary output, which is inadequate to guide evolution. By retrieving information provided by intermediate results of the classification task, it became possible to develop a suitable fitness function. The experimental results show the ability of the system to evolve images that are classified as faces. A subjective analysis also reveals the unexpected nature and artistic potential of the evolved images.
Chapter
Full-text available
A novel approach to the production of evolutionary art is presented. This approach is based on the promotion of an arms race between an adaptive classifier and an evolutionary computation system. An artificial neural network is trained to discriminate among images previously created by the evolutionary engine and famous paintings. Once training is over, evolutionary computation is used to generate images that the neural network classifies as paintings. The images created throughout the evolutionary run are added to the training set and the process is repeated. This iterative process leads to the refinement of the classifier and forces the evolutionary algorithm to explore new paths. The experimental results attained across iterations are presented and analyzed. Validation tests were conducted in order to assess the changes induced by the refinement of the classifier and to identify the types of images that are difficult to classify. Taken as a whole, the experimental results show the soundness and potential of the proposed approach.
Conference Paper
Full-text available
This paper describes our first trial to build a prototype of support system for font design utilizing the technique of Interactive Evolutionary Computation. The target domain is Japanese Katakana constructed from very simple stroke elements. Some parameters for drawing elements are encoded on the genome of each candidate individual. Starting from the initial population of sixteen individuals with random genes, the user can breed his/her favourite fonts through the graphical user interface based on a framework of Simulated Breeding. Each candidate is shown in the sub-window with a sample word arbitrarily set up by the user. We certified that it effectively works to produce a type of expressive glyphs.
Conference Paper
Face detection using local successive mean quantization transform (SMQT) features and the sparse network of winnows (SNoW) classifier has received interest in the computer vision community due to its success under varying illumination conditions. Recent work has also demonstrated the effectiveness of this classification technique for character recognition tasks. However, heavy storage requirements of the SNoW classifier necessitate the development of efficient techniques to reduce storage and computational requirements. This study shows that the SNoW classifier built with only a limited number of distinguishing SMQT features provides comparable performance to the original dense snow classifier. Initial results using the well-known CMU-MIT facial image database and a private character database are used to demonstrate the effectiveness of the proposed method.
Article
In 1991, Karl Sims presented work on artificial evolution in which he used genetic algorithms to evolve complex structures for use in computer generated images and animations. The evolution of the computer generated images progressed from simple, randomly generated shapes to interesting images which the users interactively created. The evolution advanced under the constant guidance and supervision of the user. This paper describes attempts to automate the process of image evolution through the use of artificial neural networks. The central objective of this study is to learn the user's preferences, and to apply this knowledge to evolve aesthetically pleasing images which are similar to those evolved through interactive sessions with the user. This paper presents a detailed analysis of both the shortcomings and successes encountered in the use of five artificial neural network architectures. Further possibilities for improving the performance of a fully automated system are also discussed.