Master of Science in Computing: AI Art Generation Project - Fall 2024

Desklib provides past papers and solved assignments for students. This project explores AI-generated art.

Project Title: Develop Art paintings by looking at particular art (Working project: You say I
paint)
Course: Master of science in computing
Introduction
For research projects
The motivation of this work comes with face recognition and image processing techniques. The
techniques used in this can generate artistic images by machine with help of various techniques
as by artificial intelligence and machine learning. There various techniques which can be used as
convolutional neural networks (CNN), Generative Adversarial networks (GAN), keras,
Programming with help of python, etc. The various styles can be determined by use of these
algorithms.
The research question arises on difference between machine perception and human perception.
How a human can perceive image and build painting and how a machine can generate painting
with help of few parameters. How their perceptiveness can be made better in quality?
The answer can be given by wonderful technologies of machine learning or different soft
computing techniques. The techniques through which artistic images can be generated and style
can be modified with different parameters. The techniques used are as CNN (Convolutional
neural networks), Programming by python, GAN (Generative Adversarial Networks), keras, etc.
The programming can be done through sorting.
For Development projects
The domain is image recognition. The image is recognized and art can be generated. The simple
scene descriptions can be generated by different styles and can learn different style parameters.
There are various literatures discussed on these topics.
To perceive by images and building quality artistic images is one of the prioritized elements. The
styles can be understood with various informative data. The historical images and statues can be
easily understood and clarified by machines. The modern world is growing with artificial
intelligence. So, history also needs help of artificial intelligence to dig out the important insights
that can be useful for future.
The design can be done by various artificial intelligence techniques. In artificial intelligence,
there are major fields such as machine learning, soft computing techniques, robotics, etc. In
machine learning, there are various techniques as Programming with help of R, Python, etc.,
Keras, CNN, etc.

Literature Review
There are various techniques which can be used to generate art and different style related to it.
These are:
1. Programming: There are various programming languages such as Python, R, Java, etc. In
this sorting techniques can be applied to generate artistic paintings and on different style
parameters. There are various sorting techniques applied as insertion sort, Shell sort, etc. This
can be applied with different RGB colors. This can be sorted efficiently.
Figure 1: Random values in RGB
(Source: makeartwithpython)
The sorting can be done by HSV. It means Hue, Saturation and value. By implementing various
techniques of sorting the colors can be saturated and painting can be made. These techniques are
Bubble sort, Heap sort, and Quick sort. Other types of sort can also be applied. The below figure
can be seen while these types of sorting (Kaiser, 2017).
Figure 2: Sorting done in RGB Colors
(Source: makeartwithpython)

2. Neural Networks: The below image is made with the help of neural network.
Figure 3: Image made by neural network
The creative process is changing by step towards artificial intelligence. There is various system
to generate art. This can be seen in below figure.
Figure 4: CNN flow of work
Through CNN deep learning is applied. This can be done by analyzing visual images. This is
what that can be done through deep learning (Romero Martinez, 2018).

Figure 5: Painting by applying the CNN method
(Source: Medium)

Researcher
Code as Art-Art as Code - paintings
In this literature, there are various parameters which were focused on. There were as poems and
paintings based on code as art and Poems and paintings based on Art as code. The focus for this
literature would be Code as art on Paintings and art as code on Paintings. Programming is
generally conceptualized on real word scenarios. So, for modeling, all real-world scenarios
paintings could be made by using various programming technologies. The setup was done for
various groups to make various pieces of paintings. Some groups have implemented insertion
sort and then programming for that is performed. The algorithm or style of programming would
be dependent on image. The other group implements sorting in dada where no rules are there.
And some of them implemented through shell sort (Hermans, 2017).
Recognizing Art Style with deep learning.
Through artistic movement or artistic style, both historical data and visual data can be collected
through it. The identification of artistic style is crucial for indexing huge databases. Deep
residual network is used to resolve this issue. They get an accuracy of 62 %. The dataset as
wikipaintings dataset is used. They have covered 25 different styles. For this process, the training
was done by ImageNet. For providing best performance, 20 different layers are retrained to get
more effective and efficient results. The different features from image are collected and
retrained. To increase accuracy bagging is introduced in process. This will ensure best accuracy
results. Different weights are introduced by keras framework. The program was based on python.
For this keras framework is used. TensorFlow is used for back end. Here the convolutional
neural networks are used to detect results. Deep learning provides efficient and effective results
with 62 % of accuracy (Lecoutre, Negrevergne, & Yger, 2017).
Neural algorithm of artistic style
There are various capabilities, artificial intelligence system is not used. There are various
methods for which deep neural networks are used. For separation of images and re-joining them
the neural networks are used. The forward path is used for preciseness of image. In this literature
also convolutional neural networks are focused which goes in feed-forward manner. Various
features as weights are provided to the neural system process. The output comes as filtered
versions of this. Through this representation of images can be there.

Figure 6: CNN (Convolutional Neural Network)
By above figure we can understand how style and content into it can be represented. CNN are
separable in nature which is main findings of this paper. The manipulations can be performed on
different representation of images (Gatys, Ecker, & Bethge, 2015).

Art and Texture with deep neural networks
The visual information can be retrieved by computer systems from different angles. By using
texture synthesis and CNN, Neuroscience can manipulate different images and vision of that
image or painting. The painting or visual appearance is almost same. So, by this, some spatial
information is not reached. The texture can be formulized mathematically and results can be
carried out. Through modern revolution, greater steps in machine learning and artificial
intelligence make impossible imagination to possible.
Figure 7: One image is transformed to several other types of images by deep learning.
I above image we can see how a system transform the look and feel of the image. By visual
neuroscience, there are great steps towards computer vision community. The best and effective
models for managing these visual images are through by CNN and VGG networks (Gatys, Ecker,
& Bethge, 2017).
Generating art by styles and style networks
In this literature, a new system is proposed for generation of artistic visuals. The art is generated
by perceiving its style and arousing potential is increased. This is possible through Generative
Adversarial Networks (GAN). To generate creativity networks are limited in nature. The ability
of machine can be measured by human beings. The model proposed get the better results but not
better than CNN networks. The series of images can be seen in below figure (Elgammal, et al.,
2017)

Figure 8: Images generated by CNN

Development Projects
There are various papers which are related to this topic which are as follows:
In the first paper, through programming, the artistic paintings are generated. This is done with
help of sorting algorithm. This can be programmed in various languages like Java, Python, R
programming language etc. This can be done through various sorting algorithms. These can be
Quick sort, Insertion sort, heap sort, Bubble sort. Sorting can be applied to RGB Colors
(Hermans, 2017).
According to second paper, through artistic movement or artistic style, both historical data and
visual data can be collected through it. The identification of artistic style can be done by
indexing huge databases. Deep residual network is used to resolve this issue. They get an
accuracy of 62 %. Wikipaintings dataset is considered and incorporated into process. They have
covered 25 different styles. In this keras framework and python is effectively used. CNN
networks are applied (Lecoutre, Negrevergne, & Yger, 2017).
According to third paper, there are various methodologies for which deep neural networks are
used. For separation of images and re-joining them the neural networks are used. The forward
path is used for preciseness of image. Here, convolutional neural networks are focused which
goes in feed-forward manner. The output is in filtered ways (Gatys, Ecker, & Bethge, 2015).
In forth paper, the visual information can be retrieved by machines from different angles. By
using texture synthesis and CNN, Neuroscience can manipulate different images and vision of
that image or painting (Gatys, Ecker, & Bethge, 2017).
In fifth paper, the art is generated by perceiving its style and arousing potential is increased. Here
Generative Adversarial Networks (GAN) is proposed. To generate creativity networks are
limited in nature. The ability of machine can be measured by human beings. The model proposed
get the better results but not better than CNN networks (Elgammal, et al., 2017).

Project Management
For this, various literatures and methods are studied. Both programming and machine learning
techniques are used. The programming can be done by python, java or R programming language.
Deep learning techniques are used such as CNN, VGG and GAN Networks. Through these
techniques, artistic images can be generated and style according to different parameters. For
styling different sorting algorithms are used.
Figure 9: Giant chart
The above figure is giant chart made for this work.

Ethical issues:
o Increasing the accuracy of system: The accuracy can be increased by implementing various
techniques. One of the techniques is bagging. It is implemented in one of papers to increase
accuracy of that particular model.
o Determining various parameters: The parameters taken to implement the model should be
correctly placed and identified to increase the productivity of model (Radu, 2017).
o For processing, the data set should be huge. Otherwise, new instances would not be clearly
identified.
o The error rate should be determined for the productivity of particular model.
If programming is used to generate art then python framework should be there with its database.
If java is used then its database would also be required. If R programming is used then its
framework with other libraries and dataset is required.
The neural networks can be implemented by using python or R programming language. Different
modules and libraries are needed to be loaded. The dataset containing these images and its
crucial information with metadata should also be incorporated in files. All the requirements
needed are easily available.

Ethics
Ethical issues:
o Increasing the accuracy of system: The accuracy can be increased by implementing various
techniques. One of the techniques is bagging. It is implemented in one of papers to increase
accuracy of that particular model.
o Determining various parameters: The parameters taken to implement the model should be
correctly placed and identified to increase the productivity of model (Radu, 2017).
o For processing, the data set should be huge. Otherwise, new instances would not be clearly
identified.
o The error rate should be determined for the productivity of particular model.
o The results sometimes must not be accurate. But they are somewhat closer to the results.
o Heuristics are needed to be applied to determine various parameters while implementing
neural networks.
o Exception handling should be used while programming I any of the languages such as
python, R programming language, etc.
o If there are many parameters then it should be reduced by using the concept of principal
component analysis (PCA).
o Biasness in model should be correctly identified to yield better results (cai, 2017).