=Paper=
{{Paper
|id=Vol-2744/short42
|storemode=property
|title=Ergonomic Support for Logo Development Based on Deep Learning (short paper)
|pdfUrl=https://ceur-ws.org/Vol-2744/short42.pdf
|volume=Vol-2744
|authors=Alexandr Kuzmenko,Sergey Kondratenko,Konstantin Dergachev,Valery Spasennikov
}}
==Ergonomic Support for Logo Development Based on Deep Learning (short paper)==
https://ceur-ws.org/Vol-2744/short42.pdf
Ergonomic Support for Logo Development Based on
Deep Learning*
Alexandr Kuzmenko, Sergey Kondratenko, Konstantin Dergachev, and
Valery Spasennikov
Bryansk State Technical University, Bryansk, Russia
alex-rf-32@yandex.ru
Abstract. Every year rendering logos becomes an increasingly important task
in various fields. One of the most interesting methods for rendering logos is the
use of neural networks. This paper proposes a method for rendering logos using
a convolutional neural network (CNN), specially trained to classify objects
based on a single keyword and to select parametric characteristics of the logo.
Special attention is paid to the ergonomic evaluation of resulting logos and the
feasibility of the proposed method is experimentally confirmed. The research
has shown that the results obtained are superior compared to the most modern
approaches.
Keywords: Ergonomics, Design, Logo, Vector Graphics, Expert Analysis, Us-
ability Testing, Color, Font, Shape, Visualization, Convolutional Neural Net-
work, Deep Learning.
1 Introduction
Logos, also known as trademarks, are important in today's marketing world. Logo
rendering is a key issue in a wide range of areas.
Today, the logo is one of the best tools for illustrating what a commercial organiza-
tion does, what its nature, politics and purposes are. In fact, professional logo design
provides its recognition and an organization can build its brand on it.
As a rule, the most discussed logos are aesthetically appealing, distinctive, memo-
rable, scalable, easy to use, adaptable (in color and black and white), and they effec-
tively convey the characteristics of the organization. Based on the mentioned above, it
can be argued that creating an effective visual representation of the brand requires
much more than just graphic design. For this reason, the paper emphasizes the rules
that were laid down in the basis of functioning the developed system [1].
1. Balance. Balance is important in logo design, because the brain perceives a bal-
anced design as pleasant and attractive. One can achieve balance by maintaining
the" weight " of graphics, color, size, and symmetry.
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons Li-
cense Attribution 4.0 International (CC BY 4.0).
2 A. Kuzmenko et al.
2. Size. A logo is not effective if it loses too much clarity when reduced or increased
in size in the case of being applied to printing and other advertising materials.
3. Sequence of color palettes. When using colors in the logo design, one should main-
tain a constant color palette.
When operating with colors while rendering logos a number of rules should be
considered:
1. Use colors that are additional and similar.
2. Make sure that the logo looks good in a grayscale, black and white, and a two-
color palette.
3. Select colors based on the color circle.
4. Use fonts. Using fonts allows to add identity to the created logo.
5. Form. Using simple forms makes it as easy as possible to perceive and remember
the logo.
2 Methods of system operation
The logo rendering method is based on sampling the output space of bounding boxes
into a default set, depending on the location of the map of objects [2]. During render-
ing the network generates ratings for each logo element in each box and makes ad-
justments for the box according to the specified parameters to design and visualize the
shape of the object better. In addition, the network combines forecasts from a variety
of characteristic charts for natural rendering of various logo sizes. Experimental re-
sults for data sets confirm that the developed network has competitive accuracy, while
providing a unified structure for learning and output [3,4].
At the moment, almost no work has been found on rendering logos using ANN
(artificial neural networks). Most of works are dedicated to recognizing logos on im-
ages and in video streams rather than rendering them. [5] Fig. 1 shows the architec-
ture of one of the logo recognition networks, which is based on the multibox concept.
Fig. 1. Model architecture
3 Logo rendering process
Logos of garden centers were selected for training the neural network. The total num-
ber of logos is 1582. As the next step, the images were annotated. To do this, we used
Ergonomic Support for Logo Development Based on Deep Learning 3
LabelImg, a tool for annotating graphic images written in Python, which uses Qt for
its own graphical interface. The created annotations are saved as XML files in
PASCAL VOC format used by ImageNet. It also supports YOLO format [6].
When working with large data sets, using a binary file format for storage can have
a significant impact on performance, import, and, as a result, on the learning time of
our model. Binary data takes up less disk space, less time is spent on copying it, and
they can be read more efficiently. TFRecord file formats were selected for operation.
They allow to combine multiple data sets easily and are integrated with the import
and preprocessing functions provided by Tensorflow library [7].
All the objects on the logos of competitors were classified. At the next stage, the
neural network was trained to select an object on the basis of a database of graphic
icons that can be added to the created logo. In the general case the following classes
of objects were obtained:
─ Harvest
─ Equipment
─ Vegetables
─ Fruits
─ Flowers
─ Font elements
4 Converting XML into TFRecord
To convert XML files into TFRecord, they are first converted into CSV. Typically,
XML files are converted into two CSV files, one for the final version and one for the
test. After XML files have been converted into CSV files, they are converted into
TFRecords using Python script.
To render a logo based on pre-defined characteristics, transfer learning can be used
if it is necessary to study a new object. The advantage of transfer learning is that
learning can be faster, and the required data that may be needed is not as large. Ten-
sorFlow has quite a few pre-trained models with available checkpoint files as well as
configuration files [8].
The last thing that is necessary to do before starting training is create a placemark
map. The placemark map is basically a dictionary that contains the class ID and name.
5 Logo rendering
To check how well the model works, it is necessary to write a word describing the
main activity of the organization in the input field. This paper deals with an example
of "garden" direction and botanysad.ru keyword (Fig. 2). The main parameters of the
logo are:
─ Color scheme (rgb)
─ Brightness
4 A. Kuzmenko et al.
─ Contrast ratio
─ Shape
─ Symmetry/asymmetry
─ Size
─ Type
Fig. 2. Examples of the obtained logos
To evaluate the result from the point of view of ergonomics and applicability, the
expert evaluation method was used [9].
Given the requirements for the competence of experts, as well as organizational
factors, it is not entirely appropriate to involve a large number of experts to evaluate
the results of the proposed neural network. It is optimal to involve 10 to 15 experts for
this task. On the one hand, such a number may already show a certain range of opin-
ions, on the other hand, the organization and processing of the results of such research
will be relevant to the scale of the task.
At the same time, the members of the selected expert group should be paid atten-
tion to. It is necessary to attract experts with different competencies – they can be
ergonomists, designers, and representatives of the target audience.
Experts are invited to evaluate the logo for compliance with the following criteria
[10]:
1. Scalability is one of the most important ergonomic requirements for logos, which
characterizes the possibility of using it. The logo should look complete in different
formats and sizes.
2. Producibility is the criterion that evaluates the possibility of reproducing the logo
on different surfaces and using different printing technologies.
3. Succinctness is one of the first perception criteria, which is responsible for the
harmonious combination of elements.
4. Aesthetics is a fairly subjective criterion that assesses the overall perception of the
logo.
5. Memorability is the most important criterion for the logo success. In case of suc-
cessful implementation, it turns out to create a close relationship between the im-
age and the brand.
6. Protective power determines the possibility of using the logo for brands from a par-
ticular area. A logo can be recognized as non-protectable based on the criteria of
Ergonomic Support for Logo Development Based on Deep Learning 5
protective power provided in Article 6 of the Law of the Russian Federation "On
trademarks, service marks and appellations of goods origin".
7. Unique character is one of the most important criteria for evaluating a logo. Only a
unique logo can be registered and protected. In Russia, this is done by the Federal
Service for Intellectual Property, Patents and Trademarks.
8. Associativity is a fairly subjective criterion that assesses how well the brand and its
symbolic image in the form of a logo correspond to each other. What associations
it causes in the target audience. One of the most difficult criteria in terms of auto-
mated assessment.
9. Relevance. This criterion is used to evaluate the compliance of the logo's applica-
tion area and its visual image.
Experts are asked to evaluate whether the logo meets or does not meet the above
criteria in binary format.
Various methods can be used for statistical processing of expert evaluations. As a
part of this work, expert evaluations themselves are not the subject of research, but
only a tool for evaluating the results of the developed neural network, the first im-
portant criterion for evaluating expert responses is their consistency. The most com-
mon way to assess the consistency of expert opinions is Kendall’s rank correlation
coefficient. Taking into account mentioned above, we decided to focus on evaluating
the consistency of expert opinions using Kendall’s rank correlation coefficient with-
out using more complex tools from the apparatus of reducing the number of variables
and methods of multidimensional data analysis. Concordance coefficients W for the
corresponding groups of logo requirements are calculated using the following formu-
la:
where n is the number of respondents (experts), m is the number of parameters, ac-
cording to which the evaluation is made, ry is the total evaluation of the logo by n-
expert.
To evaluate the statistical significance of the coefficient of evaluation consistency
W, we should calculate the inverse one-way probability of the distribution that is
Х2= m*(n-1)*W. The calculated indicator Х2 can be obtained using MS Excel table
editor using Х2ОБР function. The actual indicator should be higher than the calculat-
ed one, which will indicate that the concordance coefficient is statistically significant.
The logo proposed by the neural network was evaluated by a pre-selected expert
group. Based on the results, the final evaluation (compliance/non-compliance) was set
for each criterion and the concordance coefficient was calculated for each of the crite-
ria (Table) [11].
6 A. Kuzmenko et al.
Table 1. Expert evaluation of the logo for compliance with the logo main criteria.
Logo evaluation criterion Expert Concordance coefficient
evaluation
Scalability + 0,97
Producibility + 0,87
Succinctness + 0,67
Aesthetics + 0,67
Memorability + 0,78
Protective power + 0,67
Unique character - 0,67
Associativity + 0,87
Relevance + 0,67
It should be noted that expert evaluations are quite consistent, which speaks in fa-
vor of their statistical significance. From the point of view of further development of
the algorithm, it is worth paying attention to such criteria as protective power and
unique character. These requirements for the logo are necessary to take into account
additionally. It is possible to connect a wide database of registered trademarks and
logos and requirements for their registration.
6 Conclusion
The proposed model for developing logos and brand names using a convolutional
neural network and deep learning methods makes it possible to simplify and reduce
the cost of developing various logo variants significantly. Using this model, it is pos-
sible to generate a large number of different variants of logos, then on the basis of
expert evaluation method and usability testing to choose the best option and, if neces-
sary, to evolve it using classic computer graphics tools. This approach allows to de-
velop quickly unique logos for a variety of brands in the case of appropriate training
of the neural network.
At this stage, the developed information system allows to render fairly simple log-
os. The service takes into account the main ergonomic requirements for creating logos
and modern approaches of designers to drawing logos. The existing services for au-
tomatic logo rendering are inferior to the developed system in the following indica-
tors:
1. The number of independent unique images that use the same elements. The system
does not create logo repetitions.
2. Combining color schemes based on a color circle.
3. Construction of a logo of any size.
Ergonomic Support for Logo Development Based on Deep Learning 7
References
1. Mikhalina D.M, Kuzmenko A.A., Dergachev K.V., Vitaliy Shkaberin V.A. Image Colori-
zation. GraphiCon 2019. Computer Graphics and Vision. Proceedings of the 29th Interna-
tional Conference on Computer Graphics and Vision Bryansk, Russia, September 23-26,
2019. P. 207-210
2. Requirements to logos. Logo Classification. Design. Computer Graphics as Art. URL:
http://d2-art.jimdo.com (accessed 15.05.2016).
3. Averchenkov V.I., Kondratenko S.V., Spasennikov V.V. The application of the scale of
individual color preferences of respondents in testing methods. Information Systems and
Technologies, 2016, no. 2 (94). Pp. 5-13.
4. Liang, Xiangguo, et al. «Deep patch-wise colorization model for grayscale images»
SIGGRAPH ASIA 2016 Technical Briefs. ACM, 2016.
5. Cheng, Zezhou, Qingxiong Yang, and Bin Sheng. «Deep colorization» Proceedings of the
IEEE International Conference on Computer Vision. 2015.
6. Goodfellow, Ian, et al. «Generative adversarial nets» Advances in neural information pro-
cessing systems. 2014.
7. Stolbova I.D., Aleksandrova E.P., Nosov K.G. Geometric modeling as a component of
computer design . Design. Theory and Practice, 2014, no.17. Pp. 61-75.
8. Averchenkov V.I., Gulakov V.K., Mirochnikov V.V., Potapov I.A., Spasennikov V.V.,
Trubakov
9. A.O. Formation of the color palette for content based image retrieval automated systems //
World applied sciences journal. - 2013. - Т. 24. № 24. - С. 1-6.
10. Zhang, Richard, Phillip Isola, and Alexei A. Efros. «Colorful image colorization» Europe-
an Conference on Computer Vision. Springer International Publishing, 2016.
11. 6. Medsker, L. R., and L. C. Jain. «Recurrent neural networks» Design and Applications 5
(2001).
12. Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Sala-
khutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of
machine learning research, 15(1): 1929–1958, 2014.